refactor(cuda): adapt RearrangeInfinilm to Copy

[voltjia]

Summary

• Implement the deprecated CUDA-family RearrangeInfinilm provider as a thin adapter over canonical Copy.
• Remove the duplicate metadata allocation, launcher, and direct kernel dispatch while preserving the legacy base interface and tests.

Motivation

RearrangeInfinilm(input, out) is the non_blocking=false subset of the existing Copy(src, non_blocking, out) operator. Keeping two provider implementations would duplicate the same strided copy behavior during the InfiniLM compatibility migration.

No issue is closed by this PR.

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

NVIDIA validation used accelerator-dev/nvidia:latest on an A100 host and the current InfiniRT install prefix:

cmake -S /workspace/InfiniOps -B /workspace/InfiniOps/build-adapters \
-G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_CUDA_ARCHITECTURES=80 \
-DWITH_NVIDIA=ON -DGENERATE_PYTHON_BINDINGS=ON \
-DINFINI_RT_ROOT=/opt/infinirt \
-DINFINI_OPS_OPS=copy,rearrange_infinilm \
-DFETCHCONTENT_SOURCE_DIR_CUTLASS=/tmp/cutlass-src
cmake --build /workspace/InfiniOps/build-adapters --target ops --parallel 8
python3 -m pytest tests/test_copy.py tests/test_rearrange_infinilm.py \
--devices nvidia -q
54 passed in 2.39s

Repository checks:

clang-format 21.1.8 --dry-run --Werror: passed
ruff 0.15.22 check: passed
ruff 0.15.22 format --check: 148 files already formatted
python -X utf8 scripts/generate_wrappers.py \
--devices nvidia iluvatar metax moore \
--ops copy rearrange_infinilm --strict-ops
passed

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	focused ops build passed	54 focused tests passed on A100
Iluvatar	Yes	strict wrapper generation passed	Native hardware not available; platform CI/owner validation required
MetaX	Yes	strict wrapper generation passed	Native hardware not available; platform CI/owner validation required
Cambricon	No	N/A - not affected	N/A - not affected
Moore	Yes	strict wrapper generation passed	Native hardware not available; platform CI/owner validation required
Ascend	No	N/A - not affected	N/A - not affected

Full focused `pytest` output

Running 54 items in this shard
...................................................... [100%]
54 passed in 2.39s

Benchmark / Performance Impact

N/A. The adapter calls the same canonical copy kernel and this PR makes no performance claim.

Notes for Reviewers

Alignment table

Legacy interface	Canonical interface	Open-source target
RearrangeInfinilm(input, out)	Copy(src=input, non_blocking=false, out)	PyTorch Tensor.copy_(src, non_blocking=False)

• InfiniOps retains its explicit out form so the output allocation remains caller-owned.
• The legacy class, public signature, backend registrations, and tests remain in place and remain deprecated.
• The adapter forwards the selected stream to the cached CudaCopy provider before every call.