[TLE][Ascend] Add minimal INT8 GM-to-L1 ND2NZ primitive

[ggbondbest]

Add data_copy_gm_to_l1_nd2nz_int8, a TLE custom primitive for one signed INT8 GM-to-L1 ND→NZ transfer. It follows CANN 9.1's DataCopyGM2L1ND2NZImplBase and preserves all eight Nd2NzParams fields and their uint16_t types. C++ contains one copy_gm_to_cbuf_multi_nd2nz_b8 call; allocation, tiling, matrix multiplication, scaling and scheduling stay in the caller.

• Register/export the primitive, include normal/mixed-core bitcode builds, and document the tl.dot consumer requirement.
• Validate static parameter bounds and output capacity, including a static destination bound with dynamic source strides; reject Python booleans to avoid an i1/uint16_t ABI mismatch.
• Add exact identity-matmul tests with random INT8 data, padded row strides, offsets and output guards, plus an optional equivalent-TLE A/B benchmark.

Validation on Ascend910B4-1 / CANN 9.1: all changed-file pre-commit checks, 3 valid and 33 rejected registration signatures, 6 device cases, and both manual and Ninja CMake custom-op bitcode builds passed. This builds the custom-op target, not the full compiler. Device coverage is nd_num=1; dynamic parameters have registration-only coverage. CANN 9.1 requires enable_legacy_insert_load_store_for_mix_cv=True for the existing custom-MTE2 output-memory-scope inference issue.

In the FlagGems-vLLM INT8 einsum consumer, switching only the input-copy path under identical compiler options gave median TLE/custom speedups of 1.307× (Flash B=128) and 1.095× (Pro B=4096), with per-call activation packing included and four paired blocks per shape. BF16 outputs matched exactly; FP32 checks passed. The compiler uses 4× temporary workspace for custom (256/512 KiB versus 64/128 KiB). Consumer PR: flagos-ai/FlagGems-vllm#894.

[CLAassistant]

All committers have signed the CLA.