Add as_strided metadata views

[voltjia]

Summary

This change registers aten::as_strided for the PrivateUse1 backend so PyTorch composite view operations such as matrix transpose can create Infini tensors that share their source storage.

The kernel follows PyTorch's native TensorImpl::VIEW construction and uses at::native::setStrided for symbolic metadata validation, negative-stride rejection, storage-offset handling, and storage-bound checks. It performs no data copy and provides no CPU fallback.

The conformance coverage compares matrix transpose against the CPU numerical oracle and verifies explicit and inherited storage offsets, storage aliasing, invalid metadata behavior, and source-storage lifetime. This PR is stacked on #36 because the lifetime behavior depends on the storage ownership established there.

Testing

• Ruff 0.15.21 and clang-format 19.1.4 checks passed.
• An installed wheel built against PyTorch 2.12.0+cpu passed the 8 focused tests and the full suite with 170 passed, 42 skipped, and 2 xfailed.
• An installed wheel built against PyTorch 2.13.0+cpu passed the 8 focused tests and the full suite with 170 passed, 42 skipped, and 2 xfailed.
• An installed wheel built against PyTorch 2.13.0+cu130 passed all 8 focused tests on an NVIDIA A100-SXM4-80GB with driver 580.105.08 and CUDA toolkit 13.1.80.
• The NVIDIA InfiniRT and InfiniOps dependencies were rebuilt at commits 95c70080f9551e61241110497d163dfcdf9dc7e7 and 296271487beb594a248fd463e5fff14f7ab74293 with CUDA architecture 80.