fix: support InfiniRT metadata views

[voltjia]

Summary

• Adapt shared operator metadata handling to InfiniRT borrowed shape and stride views.
• Preserve owning metadata only where values outlive their source or require mutation.
• Convert Python tensor metadata directly into Tensor::Shape and Tensor::Strides, and allow the ATen bridge to consume generic ranges.
• Keep generated base headers synchronized with scripts/generate_torch_ops.py and extend installed C++ consumer regression coverage.

Motivation

InfiniTensor/InfiniRT#38 changes lvalue TensorView::shape() and strides() accessors to borrowed metadata views and replaces exact std::vector aliases with inline-capacity owning containers. Existing InfiniOps assumptions about exact std::vector types fail at mutable, conditional-expression, pybind, and ATen boundaries. This PR keeps transient paths borrowed and makes ownership explicit only where required.

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

InfiniRT #38 @ 1824269fc72a69d9a6276a28af41034661960588
Release CPU + NVIDIA build/install: passed
Generated ATen build: 492 ops / 643 overloads, wheel built successfully
pytest tests/test_generate_torch_ops.py -q: 12 passed
pytest tests/test_cpp_api.py -q: 5 passed
pytest tests/test_silu_and_mul.py -m smoke -q --devices cpu nvidia: 4 passed, 44 deselected
pytest tests -m smoke -q --devices cpu nvidia --ignore=tests/test_cpp_api.py: 101 passed, 8 skipped, 22328 deselected
Direct rank-8/rank-9 add on CPU and CUDA: passed

InfiniRT current master @ b548ff5a1475a3945009b101c1aadc13306ae32c
Release CPU + NVIDIA build/install: passed
CPU + NVIDIA + Torch smoke wheel: built successfully
pytest tests/test_generate_torch_ops.py -q: 12 passed
pytest tests/test_cpp_api.py -q: 5 passed
pytest tests/test_silu_and_mul.py -m smoke -q --devices cpu nvidia: 4 passed, 44 deselected
pytest tests -m smoke -q --devices cpu nvidia --ignore=tests/test_cpp_api.py: 93 passed, 16 skipped, 12965 deselected
Direct rank-8/rank-9 add on CPU and CUDA: passed

Ruff 0.15.22 check + format --check: passed
clang-format 21.1.8 dry-run: passed
git diff --check: passed

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
CPU	Yes	smoke passed with both InfiniRT revisions	Full suite not run; installed C++ and rank-8/rank-9 checks passed
NVIDIA	Yes	smoke passed with both InfiniRT revisions	Full generated ATen build passed with the exclusion noted below; full pytest not run
Iluvatar	Yes	not run	Hardware/toolchain unavailable; reviewer routing requested below
MetaX	Yes	not run	Hardware/toolchain unavailable; reviewer routing requested below
Cambricon	Yes	not run	Hardware/toolchain unavailable; reviewer routing requested below
Moore	Yes	not run	Hardware/toolchain unavailable; reviewer routing requested below
Ascend	Yes	not run	Hardware/toolchain unavailable; reviewer routing requested below
Hygon	Yes	not run	Hardware/toolchain unavailable; reviewer routing requested below

Benchmark / Performance Impact

N/A - compatibility fix. Borrowed views remain borrowed on transient paths, explicit copies are limited to owned or mutable metadata, and the Python conversion avoids an intermediate std::vector.

Notes for Reviewers

• This is a Draft because the shared wrapper changes still need platform-owner validation on Iluvatar, MetaX, Cambricon, Moore, Ascend, and Hygon. @InfiniTensor/infiniops-core-reviewers, please route those checks to reviewers with access.
• The branch is compatible with both current InfiniRT master and the metadata-view implementation from InfiniRT refactor: make data type mappings and shared CUDA headers device-aware #38.
• InfiniRT refactor: make data type mappings and shared CUDA headers device-aware #38 changes TensorView layout, so matching InfiniRT headers and libinfinirt.so must be rebuilt and deployed together.
• The unrestricted all-ATen build reaches the pre-existing std::optional<const at::Tensor> assignment error in src/torch/ops/flash_attn_with_kvcache/flash_attn.cc. The successful generated ATen build excludes only that handwritten implementation while still compiling all 492 generated ops; this PR does not modify the unrelated failure.
• tests/test_cpp_api.py requires an installed prefix. The reported five tests were run with INFINI_OPS_INSTALL_PREFIX set for each InfiniRT revision.
• Generated base headers and scripts/generate_torch_ops.py should be reviewed together.