feat(cambricon): add reshape and cache flash provider

[baominghelly]

Summary

• Add a Cambricon native ReshapeAndCacheFlash provider in src/native/cambricon/ops/reshape_and_cache_flash/.
• Scatter token-major key/value rows into paged flash-attention caches using slot_mapping, including negative-slot skips and stride-aware NHD/HND cache layouts.
• Support FP16, BF16, and FP32 on the caller-provided Cambricon stream.

Motivation

ReshapeAndCacheFlash did not have a Cambricon native provider, so Cambricon inference paths could not populate paged KV caches through the current InfiniOps interface.

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

Targeted Cambricon configure, build, and install (this is not the repository-wide smoke build):

cmake -S /workspace/src/InfiniOps-llama-reshape-cache-flash \
  -B /workspace/build/infiniops-reshape-and-cache-flash-pr \
  -DCMAKE_BUILD_TYPE=Release \
  -DWITH_CAMBRICON=ON \
  -DGENERATE_PYTHON_BINDINGS=ON \
  -DINFINI_OPS_OPS=reshape_and_cache_flash \
  -DINFINI_RT_ROOT=/workspace/install/infinirt-llama-trace \
  -DCMAKE_INSTALL_PREFIX=/workspace/build/infiniops-reshape-and-cache-flash-pr-site/infini
cmake --build /workspace/build/infiniops-reshape-and-cache-flash-pr \
  --target ops --parallel 32
cmake --install /workspace/build/infiniops-reshape-and-cache-flash-pr

[ 50%] Built target infiniops
[100%] Built target ops

Targeted Cambricon test:

PYTHONPATH=/workspace/build/infiniops-reshape-and-cache-flash-pr-site \
  /torch/venv3/pytorch/bin/python3.10 -m pytest -q \
  tests/test_reshape_and_cache_flash.py --devices cambricon

........................                                                 [100%]
24 passed in 0.20s

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	No	N/A - not affected	N/A - not affected
Iluvatar	No	N/A - not affected	N/A - not affected
MetaX	No	N/A - not affected	N/A - not affected
Cambricon	Yes	targeted ops build passed	24 targeted tests passed; repository smoke/full suite not run
Moore	No	N/A - not affected	N/A - not affected
Ascend	No	N/A - not affected	N/A - not affected

Full `pytest` output (optional)

........................                                                 [100%]
24 passed in 0.20s

Benchmark / Performance Impact

N/A - this adds a previously unavailable Cambricon provider; no performance benchmark was run.

Notes for Reviewers

• Each MLU logical task handles one or more (token, head) pairs and writes one contiguous head_size vector into the cache slot selected by slot_mapping.
• Negative slots are skipped. The caller contract must provide in-range non-negative slots and avoid duplicate concurrent destinations.
• Tensor strides are used for token/head input addressing and block/page/head cache addressing, covering padded sources plus logical NHD and HND cache layouts.
• The current public interface only accepts kv_cache_dtype == "auto", so cache dtypes match key/value dtypes and k_scale / v_scale do not participate in the copy.
• The implementation is correctness-oriented direct device copying; NRAM/vectorized performance optimization is intentionally deferred.
• Existing cross-platform tests already cover the provider, so this PR adds no test-only changes.
• The repository smoke build and full suite were not run; the PR remains draft pending broader validation.