feat(ops): add vLLM-aligned reshape_and_cache_flash

[voltjia]

Summary

• Add the public ReshapeAndCacheFlash operator and Python binding surface aligned with vLLM's public reshape_and_cache_flash wrapper.
• Reuse the canonical stride-aware CUDA kernel from both the new operator and deprecated PagedCachingInfinilm, without changing or removing the legacy public API.
• Cover NHD/HND cache strides, padded source strides, scalar/per-head scales, supported input dtypes, shortened slot mappings, and negative slots.

Motivation

PagedCachingInfinilm is an InfiniLM compatibility API. This PR adds a canonical serving-framework-aligned replacement so consumers can migrate without removing the deprecated interface.

No issue is linked; this is part of the ongoing InfiniLM operator migration.

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

API Alignment

InfiniOps API	Alignment target	Upstream public API	InfiniOps C++ adaptation
reshape_and_cache_flash(key, value, slot_mapping, k_scale, v_scale, kv_cache_dtype, key_cache, value_cache) -> void	vLLM	reshape_and_cache_flash(key, value, key_cache, value_cache, slot_mapping, kv_cache_dtype, k_scale, v_scale) -> None	Preserve the relative order of logical input tensors, place the string attribute after inputs, and move writable key_cache / value_cache outputs to the end, as required by CONTRIBUTING.md.

All eight material upstream parameters are represented. This PR adds one canonical overload.

Smoke Test Result

Remote environment: ssh nvidia, image accelerator-dev/nvidia:latest.

python3 -m pip install . --force-reinstall --no-build-isolation --no-deps \
  --config-settings=cmake.define.AUTO_DETECT_DEVICES=OFF \
  --config-settings=cmake.define.AUTO_DETECT_BACKENDS=OFF \
  --config-settings=cmake.define.WITH_CPU=ON \
  --config-settings=cmake.define.WITH_NVIDIA=ON \
  --config-settings=cmake.define.INFINI_RT_ROOT=/opt/infinirt \
  --config-settings=cmake.define.INFINI_OPS_SMOKE_BUILD=ON
Successfully built InfiniOps

python3 -m pytest tests -m smoke -q
91 passed, 22 skipped, 10043 deselected in 20.35s

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	Smoke build/test passed; focused build passed	Focused new/legacy tests: 54 passed, 30 CPU-only legacy skips
Iluvatar	Yes	Not run locally; CI required	Shared generic CUDA provider
MetaX	Yes	Not run locally; CI required	Shared generic CUDA provider
Cambricon	No	N/A - not affected	N/A
Moore	Yes	Not run locally; CI required	Shared generic CUDA provider
Ascend	No	N/A - not affected	N/A

Focused test output

python3 -m pytest tests/test_reshape_and_cache_flash.py \
  tests/test_paged_caching_infinilm.py -q -rs
54 passed, 30 skipped in 5.68s

python3 -m pytest tests/test_generate_wrappers.py -q
22 passed in 2.12s

The 30 skips are the existing PagedCachingInfinilm tests on CPU, where that operator has no implementation.

Benchmark / Performance Impact

N/A. This PR adds an interface and reuses the existing cache-write kernel; it does not claim a performance change.

Notes for Reviewers

• The new public signature follows vLLM's Python wrapper, with only the repository-required input/attribute/output grouping adaptation.
• The implementation currently supports kv_cache_dtype == "auto". FP8/NVFP4 behavior can be added later without changing the public signature.
• PagedCachingInfinilm remains deprecated and source-compatible. Its launcher directly includes the canonical reshape_and_cache_flash/kernel.cuh; no forwarding header or extra compatibility layer is introduced.
• Exact formatting checks passed with clang-format==21.1.8 and ruff==0.15.22.