refactor(linked)!: migrate flash_attn_with_kvcache provider

[voltjia]

Summary

• Move the NVIDIA flash_attn_with_kvcache provider from the handwritten src/torch path to the generic src/linked/torch architecture.
• Replace the operator-specific FlashAttention discovery, ABI check, link, and RPATH CMake block with linked-library YAML metadata and the shared resolver.
• Preserve the existing public operator API and behavioral tests while assigning the linked implementation to slot 16.

Motivation

The existing integration already called the C++ ABI exported by flash_attn_2_cuda, but independently reimplemented package discovery, symbol inspection, library linking, RPATH setup, stream bridging, and ATen tensor adaptation in src/CMakeLists.txt and src/torch.

This migration makes the complex direct-symbol FlashAttention provider use the same linked-library mechanism as the simpler linked operators. It also validates that the architecture supports optional tensors, mutable caches, auxiliary outputs, scalar adaptation, and non-default streams without copying third-party kernels into InfiniOps.

This PR is stacked on #886. N/A - no linked issue.

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

Remote environment: ssh nvidia, accelerator-dev/nvidia:latest, NVIDIA A100, PyTorch 2.10.0a0+b4e4ee8, FlashAttention 2.7.4.post1+25.12.

cmake -S . -B build-linked-fa-smoke -G Ninja \
  -DINFINI_RT_ROOT=/opt/infinirt \
  -DWITH_CPU=ON -DWITH_NVIDIA=ON -DWITH_TORCH=ON -DWITH_LINKED=ON \
  -DAUTO_DETECT_DEVICES=OFF -DAUTO_DETECT_BACKENDS=OFF \
  -DGENERATE_PYTHON_BINDINGS=ON -DINFINI_OPS_SMOKE_BUILD=ON
cmake --build build-linked-fa-smoke --target ops -j 16
PYTHONPATH=build-linked-fa-smoke/install \
  python3 -m pytest tests -m smoke --devices nvidia -q

78 passed, 20 skipped, 7554 deselected in 20.60s

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	Focused linked build and default smoke passed	tests/test_flash_attn_with_kvcache.py: 12 passed
Iluvatar	No	N/A - not affected	N/A
MetaX	No	N/A - not affected	N/A
Cambricon	No	N/A - not affected	N/A
Moore	No	N/A - not affected	N/A
Ascend	No	N/A - not affected	N/A

Additional validation output

python3 scripts/resolve_linked_ops.py \
  --devices nvidia \
  --ops flash_attn_with_kvcache \
  --nm nm --readelf readelf --cxxfilt c++filt
exit 0

cmake -S . -B build-linked-fa -G Ninja \
  -DINFINI_RT_ROOT=/opt/infinirt \
  -DWITH_NVIDIA=ON -DWITH_LINKED=ON -DWITH_TORCH=OFF \
  -DGENERATE_PYTHON_BINDINGS=ON \
  -DINFINI_OPS_OPS=flash_attn_with_kvcache
cmake --build build-linked-fa --target ops -j 16

-- Resolved 1 linked operator source(s).
[build completed]

PYTHONPATH=build-linked-fa/install \
  python3 -m pytest tests/test_flash_attn_with_kvcache.py \
  --devices nvidia -q

12 passed in 2.12s

Benchmark / Performance Impact

N/A - this PR changes provider discovery and integration structure, not the called FlashAttention ABI or operator computation.

Notes for Reviewers

• The provider still calls the exact flash::mha_fwd_kvcache C++ ABI exported by the installed flash_attn_2_cuda DSO. The full demangled signature is now verified by flash_attn.yaml; a package version is no longer treated as an ABI proxy.
• The generic template owns ATen conversion, cache mutation, auxiliary LSE copy-back, scalar cache_seqlens adaptation, and external-stream guarding. The NVIDIA provider contains only the library ABI declaration/call and slot specialization.
• No third-party kernel source or generated kernel is copied into the repository.
• The current stack base predates master commit 1dc5014f (fix: copy optional tensor metadata #881). Manual validation applied that already-merged optional-metadata fix only to the remote test copy; it is not part of this PR. The branch should be rebased onto current master after its seven linked prerequisites merge.
• The existing operator test file is unchanged and covers dense and paged caches, KV append, tensor/scalar sequence lengths, LSE output, defaults, and a non-default CUDA stream.

BREAKING CHANGE: enabling the NVIDIA flash_attn_with_kvcache provider now requires WITH_LINKED=ON instead of relying on WITH_TORCH=ON, and its implementation index moves from slot 8 to linked slot 16. The public operator signature and behavior are unchanged.