C-V mix kernel example, todo items

[learning-chip]

Start C-V mix kernel example in Python, based on: hw-native-sys/PTOAS#213.

IR doc and unit tests:

• https://github.com/zhangstevenunity/PTOAS/blob/main/docs/designs/ptoas-tpush-tpop-design.md
• https://github.com/zhangstevenunity/PTOAS/tree/main/test/samples/TPushTPop
• https://github.com/zhangstevenunity/PTOAS/blob/main/test/basic/tpush_tpop_dir_both.pto
• https://github.com/zhangstevenunity/PTOAS/blob/main/test/basic/tpush_tpop_frontend_lowering_a3.pto
• More samples: test: add A3 TPushTPop sample cases hw-native-sys/PTOAS#382

Related C++ doc:

• https://github.com/PTO-ISA/pto-isa/blob/main/docs/HL_ptoisa_newfeature20260306_TPUSH_TPOP.md
• the raw ffts_cross_core_sync intrinsic on A2

Unit tests:

• https://gitcode.com/cann/pto-isa/blob/master/tests/npu/a2a3/src/st/testcase/tpushpop_cv/
• https://gitcode.com/cann/pto-isa/tree/master/tests/npu/a2a3/src/st/testcase/tpushpop_vc
• FA also uses PUSH/POP API now https://gitcode.com/cann/pto-isa/blob/master/kernels/manual/common/flash_atten/fa_performance_kernel.cpp
  • the torch interface https://gitcode.com/cann/pto-isa/pull/612

[learning-chip]

Concrete kernel lists:

• Start with simple case such as this linear attention
  • should support run-time dynamic-shape, like most pto-dsl examples, but unlike tilelang's static example
• Then generalize to chunk GDN & KDA
• Ref full list [User Story] Collection and classification of new kernel needs pto-kernels#23

[learning-chip]

Port a minimum C-V sync example like huawei-csl/pto-kernels#65 to Python DSL

• Include both C2V and V2C sync, use manual sync everywhere
• Test manual sync between cores (using PUSH/POP API) together with auto-sync within cores (ptoas insert sync)