feat(ops): add Copy operator

[voltjia]

Summary

• Adds a canonical Copy operator with the public C++ signature Copy(const Tensor src, const bool non_blocking, Tensor out).
• Adds shared CUDA implementation coverage for NVIDIA, Iluvatar, MetaX, and Moore, including broadcasted and strided source layouts.
• Keeps the deprecated RearrangeInfinilm API and tests intact while reusing CopyKernel directly.
• Adds one parameterized test_copy covering non_blocking, broadcasting, non-contiguous layouts, and float32/float16/bfloat16.

Motivation

RearrangeInfinilm was introduced as an InfiniLM compatibility operator and is now deprecated. A normal operator aligned with an established open-source API is required before downstream callers can migrate away from the suffixed interface.

API Alignment

InfiniOps API	Alignment target	Mapping	Evidence
Copy(const Tensor src, const bool non_blocking, Tensor out)	PyTorch Tensor.copy_(src, non_blocking=False) -> Tensor	src and non_blocking preserve the PyTorch order. PyTorch's mutable receiver/returned self is represented by trailing caller-provided out per the InfiniOps input-attribute-output convention. InfiniOps uses Copy/copy rather than an in-place _ suffix because mutation is expressed by out. The C++ API requires the defaulted boolean explicitly and does not add a convenience overload.	PyTorch 2.13 torch.Tensor.copy_

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

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 A100-SXM4-80GB, accelerator-dev/nvidia:latest:

python3 -m pip install . --force-reinstall --no-build-isolation --no-deps \
  --config-settings=cmake.define.INFINI_RT_ROOT=/opt/infinirt \
  --config-settings=cmake.define.AUTO_DETECT_DEVICES=OFF \
  --config-settings=cmake.define.AUTO_DETECT_BACKENDS=OFF \
  --config-settings=cmake.define.WITH_NVIDIA=ON \
  --config-settings=cmake.define.WITH_CPU=OFF \
  --config-settings=cmake.define.CMAKE_CUDA_ARCHITECTURES=80 \
  --config-settings=cmake.define.INFINI_OPS_SMOKE_BUILD=ON
Successfully built InfiniOps
Successfully installed InfiniOps-0.1.0

python3 -m pytest tests -m smoke -q
64 passed, 34 skipped, 8246 deselected in 4.13s

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	smoke passed	Focused copy + rearrange_infinilm: 54 passed, 54 skipped in 1.85s; wrapper generator: 22 passed in 1.58s
Iluvatar	Yes	Not run	Backend hardware/toolchain unavailable
MetaX	Yes	Not run	Backend hardware/toolchain unavailable
Cambricon	No	N/A	N/A
Moore	Yes	Not run	Backend hardware/toolchain unavailable
Ascend	No	N/A	N/A

Benchmark / Performance Impact

N/A. This adds an API migration path and reuses the same elementwise copy kernel for the canonical and compatibility operators.

Notes for Reviewers

• RearrangeInfinilm remains deprecated and its public signature and test coverage are unchanged.
• The current CUDA providers require matching dtype and device. PyTorch's cross-dtype and cross-device behavior is implementation work outside this interface-migration PR.
• For the currently supported same-device CUDA path, PyTorch documents that non_blocking has no effect; the parameter is retained to keep the canonical interface complete.
• No torch_ops.yaml, CMake allowlist, generated artifact, or default-value overload is added.
• Final formatting checks passed with clang-format 21.1.8 and ruff 0.15.22.