fix(tests): look for the engine where the runtime caches it

[kevin9327]

Problem

Every requires_engine test has been skipped since cd2f976, on machines that have the engine.

That commit moved the engine cache under a generation directory. needle._library_path now reads:

cache = os.path.join(os.path.expanduser("~"), ".cache", "cactus-needle",
                     f"v{generation}", version)

and needle fetch writes there, printing ~/.cache/cactus-needle/v2/2.0.4/. tests/conftest.py
was not part of that commit and still looks in the pre-split location:

cache = os.path.join(os.path.expanduser("~"), ".cache",
                     "cactus-needle", fetch.ENGINE_VERSION, name)

The v2 segment is missing, so the gate never finds a fetched engine and the six tests behind it
never run. It also misses the other two ways _library_path finds an engine that the same commit
introduced or kept: the NEEDLE2_LIB_PATH / NEEDLE_LIB_PATH override, and the generation-suffixed
in-package name (libneedle2.so).

The skip is silent, so nothing has been failing. The tests simply stopped covering anything.

Fix

_engine_available now mirrors _library_path's lookup order — override, in-package names,
generation-scoped cache — and stops short of the download that function ends with.

Tests

Two in tests/test_fetch.py, both failing before this change:

• test_engine_gate_finds_the_cache_the_runtime_loads_from — with HOME pointed at a temp tree and
  the package directory empty, the gate is false, then true once an engine file exists at the path
  _library_path uses.
• test_engine_gate_honours_the_library_override — NEEDLE_LIB_PATH at a real file opens the gate;
  pointed at a missing one it does not.

How I tested

Windows 11, Python 3.12.10, pip install -e ".[test,train]", engine fetched with needle fetch
into C:\Users\...\.cache\cactus-needle\v2\2.0.4\libneedle.dll.

The six gated tests, before and after:

before   25 passed, 6 skipped   (all six "needle C++ engine not installed")
after    31 passed

Whole suite, pytest -q -m "not slow":

before   64 passed, 9 skipped
after    72 passed, 3 skipped

The six that stop skipping are the engine tests; the two extra passes are the tests above. The three
that still skip are the worker integration tests, which want a C compiler. No test changes from
passing to failing.

Both runs exclude test_lora.py, test_render.py, test_run.py, test_build.py and
test_finetune.py, which need flax. That will not install on this machine: orbax-checkpoint
ships paths past the Windows MAX_PATH limit and pip fails with WinError 206. Nothing here
touches that path, and CI runs it on Ubuntu.

Worth noting separately: the engine tests pass on Windows once they actually run, so the gate was
hiding working coverage rather than broken code.