fix: copy optional tensor metadata

[voltjia]

Summary

• Copy optional tensor shapes into owning Tensor::Shape members instead of mixing borrowed ShapeView values with owning empty shapes.
• Copy optional tensor strides into owning Tensor::Strides members across BinaryCrossEntropy, MultiMarginLoss, and both FlashAttention wrappers.

Motivation

InfiniTensor/InfiniRT#38 changed lvalue TensorView::shape() and strides() to return borrowed metadata views. Conditional initializers that paired those views with owning empty metadata no longer had a common C++ type, so InfiniOps failed to compile against the current InfiniRT master.

The smoke-build failure was discovered by #878. A full build found the same compatibility issue in three additional consumers, which this PR fixes together.

Related:

• perf!: inline TensorView metadata storage InfiniRT#38
• fix!: align Gemm public signature #878

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

Environment:
  image: accelerator-dev/nvidia:latest
  GPU: NVIDIA A100, CUDA_VISIBLE_DEVICES=5
  InfiniRT: a38977e03f9fafd803c3a4bd0d5810c57bc5a46e
  InfiniOps: 05df7232c7864fe870353343e9d9a270e1e36a1e

$ clang-format --version
clang-format version 21.1.8

$ clang-format --dry-run --Werror <four changed headers>
FORMAT_OK

$ python3 -m pip install '.[dev]' --no-build-isolation --config-settings=cmake.define.INFINI_OPS_SMOKE_BUILD=ON --config-settings=cmake.define.INFINI_RT_ROOT=/work/infinirt-prefix
Successfully built InfiniOps
Successfully installed InfiniOps-0.1.0

$ python3 -m pytest tests/ -m smoke -n 8 -q --tb=short
114 passed, 23 skipped in 26.95s

$ python3 -m pytest tests/test_binary_cross_entropy.py tests/test_multi_margin_loss.py tests/test_flash_attn_varlen_func.py tests/test_flash_attn_with_kvcache.py -q --tb=short -rs
65 passed, 1 skipped, 8 warnings in 2.44s

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
CPU	Yes	smoke passed	Included in full build and affected-operator tests
NVIDIA	Yes	smoke passed	Full build passed; affected-operator tests passed
Iluvatar	No	N/A - not affected	N/A - not affected
MetaX	No	N/A - not affected	N/A - not affected
Cambricon	No	N/A - not affected	N/A - not affected
Moore	No	N/A - not affected	N/A - not affected
Ascend	No	N/A - not affected	N/A - not affected

The full CPU+NVIDIA+WITH_TORCH wheel built successfully against the pinned InfiniRT commit. The complete pytest suite ran, but it was not clean: all 50 failures were in unchanged test modules outside the affected operators. The pre-fix baseline cannot reach pytest against this InfiniRT commit because its full build fails on the metadata type errors.

Full build and `pytest` output (trimmed)

$ python3 -m pip install '.[dev]' --no-build-isolation --config-settings=cmake.define.INFINI_RT_ROOT=/work/infinirt-prefix
Created wheel for InfiniOps: filename=infiniops-0.1.0-cp312-cp312-linux_x86_64.whl size=37226025
Successfully built InfiniOps
Successfully installed InfiniOps-0.1.0

$ python3 -m pytest tests/ -n 8 -q --tb=short
50 failed, 13588 passed, 9161 skipped, 127 warnings in 217.93s

Failure groups:
  12  tests/test_tensordot.py
   6  tests/test_moe_sum.py
  32  tests/test_moe_wna16_gemm.py

No failure occurred in:
  tests/test_binary_cross_entropy.py
  tests/test_multi_margin_loss.py
  tests/test_flash_attn_varlen_func.py
  tests/test_flash_attn_with_kvcache.py

Benchmark / Performance Impact

N/A. This is a compile-time compatibility fix. The classes already store owning metadata; the patch makes the required view-to-owner copy explicit in conditional expressions.

Notes for Reviewers

• The owning copies are intentional: these members retain shape and stride metadata beyond the constructor expression and must not borrow from a temporary tensor view.
• A repository-wide scan found no other conditional initializer that mixes ShapeView/StridesView with owning empty metadata.
• Existing tests cover optional tensors both present and absent, so no new behavioral test case was required.