Fix the harmful uses of #define private public

[Critsium-xy]

Background

#define private public around an #include appears in 81 files in this tree
(111 occurrences). It is not a harmless test trick: it gives that translation unit
a different definition of the class than the rest of the library it links
against, which is an ODR violation, and it silently applies to every other header
pulled in by the same region.

This PR does not remove the pattern. It fixes the places where it is currently
doing damage, and removes the instances that do nothing at all.

What is fixed

1. The one occurrence in production code

source_io/module_output/cal_test.cpp is compiled into the main library
(source/source_io/CMakeLists.txt), not into a test target, and wrapped
parameter.h. The shipped binary therefore contained one translation unit that saw
Parameter::input / Parameter::sys as public while every other one saw them as
private.

The hack was also unnecessary: the file only reads PARAM.inp.nbands and
PARAM.globalv.nlocal, both public.

2. Sixteen regions that were never closed

These opened #define private public (usually protected too) and ran to end of
file — up to 1016 lines in charge_mixing_test.cpp. elecstate_pw_test.cpp was a
partial case: it closed protected 318 lines later at the end of the file and
never closed private at all.

Each #undef now sits immediately after the include block it is meant to cover.
No include is moved: in every one of these files the standard headers already came
after the project headers, so they simply fall outside the region once it is
closed. That removes the incidental treatment of <omp.h>, <string>,
<fstream>, <mpi.h>, <source_base/macros.h>,
<source_base/module_external/scalapack_connector.h> and "mpi.h" — eight files
were compiling a standard or system header with access control disabled.

Closing a region early is safe: the class definitions the tests reach into have
already been compiled by the time the #undef is seen.

3. Dead __MPI toggles, and two tests that asserted nothing

source/source_base/test/CMakeLists.txt:1 calls
abacus_disable_feature_definitions(__MPI), so __MPI is never defined for any
target in that directory and every #ifdef __MPI there is always false.

#7888 already fixed the worst of this — it removed the guards from
opt_cg_test.cpp / opt_tn_test.cpp and gave those targets a shared
mpi_test_main.cpp that actually initialises MPI. This PR clears what it left
behind:

• math_chebyshev_test.cpp — guards still wrap checkconverge and
  checkconverge_float, so those two bodies are compiled out in every
  configuration: two tests reporting PASS while asserting nothing.
• global_function_test.cpp — three #ifdef __MPI / #undef __MPI / #endif
  blocks whose test bodies sit outside the #ifdef. Pure no-ops.
• opt_cg_test.cpp, opt_test_tools.cpp — a file-level #undef __MPI that only
  duplicates what CMake already does.

The #undef __MPI / #define __MPI pairs were no-ops regardless: the preprocessor
cannot retroactively change headers already processed above, and the library these
tests link is a separate translation unit compiled with __MPI either way.

The two chebyshev bodies this exposes cannot run in MODULE_BASE_math_chebyshev —
they reach GlobalFunc::ddot_real -> Parallel_Reduce::reduce_pool, which needs a
valid POOL_WORLD, and that target has no MPI setup. They are also
byte-identical duplicates of the tests of the same names in
test_parallel/math_cheby_mpi_test.cpp (MODULE_BASE_math_chebyshev_mpi), which
sets POOL_WORLD up in its fixture and runs them properly. So they are deleted
rather than given a second MPI harness. No coverage is lost.

4. Dead code removed

• source_io/test/for_testing_input_conv.h — included by nothing. It was the only
  header applying #define private public, and applied it to 22 module headers at
  once; it also defined non-inline globals at file scope, which would be a
  duplicate symbol the moment a second TU included it.
• source_hsolver/test/test_diago_assist.cpp — in no CMakeLists.txt, includes a
  header that does not exist (diago_iter_assis.h; the real name is
  diago_iter_assist.h), and has a statement in a class body, so it cannot
  compile. Dead since [Refactor] Rename module_hsolver to source_hsolver #6305.
• Six no-op regions: unitcell_test_setupcell.cpp and unitcell_test_pw.cpp had a
  bare #define private public immediately followed by #undef private;
  write_orb_info_test.cpp, rho_io_test.cpp, deltaspin_pw_test.cpp and the
  second region of bfgs_test.cpp wrapped only parameter.h and never touch a
  private member of Parameter.

Result

	before	after
#define private public in production code	1	0
in a header	1	0
regions never closed	17	0
standard/system headers compiled inside a region	8	0
tests compiled out unnoticed	2	0
files with the pattern	81	73
occurrences	111	102

Scope

This deliberately stays narrow. It does not try to remove the pattern from the
files that wrap the header of the class under test (unitcell.h, klist.h,
charge.h, sltk_grid.h, ...) to reach its private members. Those regions are now
all balanced, tight, and free of standard headers, but the hack is still there.

In particular, 41 translation units still compile parameter.h with access control
disabled, because their tests need to write PARAM.input / PARAM.sys. An earlier
revision of this PR routed those through the TestParameters friend class that
parameter.h already declares; that was not wanted, so it has been dropped. Per
AGENTS.md rule 10 the sanctioned remedies are to promote member visibility
explicitly or add a public test-only accessor — worth settling separately, since it
touches ~1600 call sites and should be a deliberate design decision rather than a
side effect of this cleanup.

read_wf2rho_pw_test.cpp likewise keeps its file-level #undef __LCAO: it is
load-bearing (it selects the non-LCAO path in the headers below it) and belongs in
that test's CMakeLists as abacus_disable_feature_definitions(__LCAO), but
changing it changes what the test compiles.

Review notes

• Six commits, split by concern, each reviewable on its own.
• No new files and no new interfaces: the diff is deletions plus 28 inserted
  #undef lines.
• agent_governance_check.py --base upstream/develop: no blockers.
• No INPUT parameter, interface, or runtime behaviour change outside test code, so
  no docs update is required.

Testing

Verified with gcc / Ninja, BUILD_TESTING=ON,
-DENABLE_LCAO=ON -DENABLE_MPI=ON -DENABLE_OPENMP=ON. Build clean.

Unit tests compared against an upstream/develop build from the same commit base
(fb9ce1da4), same configuration, ctest run serially with OMP_NUM_THREADS=1:

	upstream/develop	this branch
tests run	335	335
failures	29	29

No regressions and no differences — the failing set is identical on both trees.
Those 29 are pre-existing in this environment: mostly mpirun-based *_para /
*_parallel wrappers, plus tests in files this PR never touches
(blas_connector, math_sphbes, cubic_spline, PSI_init, dav, bpcg).

Integration tests (tests/) were not run locally; they need the full toolchain
build, and CI covers them.

[mohanchen]

I don't like this solution.