feat(cambricon): add varlen flash attention provider

[baominghelly]

Summary

• Add the Cambricon linked provider at implementation slot 16 for flash_attn_varlen_func.
• Invoke the Cambricon flash-attn extension's mha_varlen_fwd entry point for packed variable-length attention.
• Support cumulative sequence lengths, grouped-query attention, causal/window masking, ALiBi, paged KV-cache gathering, softmax LSE output, and non-default MLU streams.
• Extend the operator tests with Cambricon references for FP16 and BF16 while preserving the NVIDIA and Moore platform-specific test behavior.

Motivation

Cambricon model prefill and variable-length batches need the FlashAttention varlen interface selected by the model operator manifest. This provider complements the KV-cache decode provider merged in #969 and reuses the linked-library declaration and paged-cache helper now available on master.

Type of Change

• feat — new feature / new operator / new platform
• fix — bug fix
• perf — performance improvement (no behavioral change)
• refactor — code restructuring without behavior change
• test — adding or fixing tests only
• docs — documentation only
• build / ci — build system or CI configuration
• chore — tooling, formatting, or other non-code changes
• Breaking change (requires a ! in the Conventional Commits prefix or a BREAKING CHANGE: footer)

Platforms Affected

• CPU (WITH_CPU)
• NVIDIA (WITH_NVIDIA)
• Iluvatar (WITH_ILUVATAR)
• MetaX (WITH_METAX)
• Cambricon (WITH_CAMBRICON)
• Moore (WITH_MOORE)
• Ascend (WITH_ASCEND)
• PyTorch C++ bindings (WITH_TORCH)
• Build system / CMake / CI
• Python bindings / user-facing API

Smoke Test Result

Not run — the repository's official Cambricon smoke suite was not run. A fresh slot-specific Cambricon linked build and operator test were run instead.

export INFINI_OPS_DISPATCH_BATCH_SIZE=64
cmake \
  -S /workspace/src/InfiniOps-cambricon-flash-attn-varlen-func-pr \
  -B /workspace/build/pr-publish-flash-attn-varlen-func-20260904-master24bf \
  -G Ninja \
  -DAUTO_DETECT_DEVICES=OFF \
  -DAUTO_DETECT_BACKENDS=OFF \
  -DWITH_CPU=OFF \
  -DWITH_CAMBRICON=ON \
  -DWITH_NVIDIA=OFF \
  -DWITH_TORCH=OFF \
  -DWITH_LINKED=ON \
  -DGENERATE_PYTHON_BINDINGS=ON \
  -DINFINI_OPS_OPS=flash_attn_varlen_func \
  -DINFINI_RT_ROOT=/workspace/install/infinirt-master-test \
  -DINFINI_OPS_TORCH_COMPILE_JOBS=2 \
  -DPython_EXECUTABLE=/torch/venv3/pytorch/bin/python3.10 \
  -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_INSTALL_PREFIX=/workspace/build/pr-publish-flash-attn-varlen-func-20260904-master24bf-site/infini
cmake --build /workspace/build/pr-publish-flash-attn-varlen-func-20260904-master24bf --parallel "$(nproc)"
cmake --install /workspace/build/pr-publish-flash-attn-varlen-func-20260904-master24bf
PYTHONPATH=/workspace/build/pr-publish-flash-attn-varlen-func-20260904-master24bf-site \
LD_LIBRARY_PATH=/workspace/build/pr-publish-flash-attn-varlen-func-20260904-master24bf-site/infini:/workspace/install/infinirt-master-test/lib \
/torch/venv3/pytorch/bin/python3.10 -m pytest -q \
  tests/test_flash_attn_varlen_func.py --devices cambricon --maxfail=1

26 passed, 5 skipped, 1 warning in 0.45s

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	No	N/A - not affected	N/A - not affected; shared test behavior preserved during rebase
Iluvatar	No	N/A - not affected	N/A - not affected
MetaX	No	N/A - not affected	N/A - not affected
Cambricon	Yes	Slot-specific fresh linked build passed; official smoke not run	26 passed, 5 skipped; one test-reference int64-to-int32 warning
Moore	No	N/A - not affected	N/A - not affected; Moore-specific skips and reference behavior preserved during rebase
Ascend	No	N/A - not affected	N/A - not affected

Targeted pytest output

26 passed, 5 skipped, 1 warning in 0.45s

Benchmark / Performance Impact

N/A - no benchmark was run.

Notes for Reviewers

• The Cambricon linked flash-attn library declaration and src/common/op_utils/paged_kv_cache.h were added by feat(cambricon): add flash attention KV-cache provider #969 and are now in master.
• The shared paged KV-cache header intentionally depends on ATen because its current consumers are linked Torch FlashAttention providers; this scope is documented inline.
• The provider was validated with dropout disabled. softcap and deterministic are currently accepted but not forwarded to the Cambricon extension.
• When return_attn_probs is enabled, the provider copies softmax LSE; the optional s_dmask output is not populated.
• Paged inputs are gathered into contiguous packed K/V tensors before calling mha_varlen_fwd.
• clang-format --dry-run --Werror and git diff --check passed.
• The official smoke suite and performance benchmarks remain validation gaps.