Preserve dense add output layouts

[voltjia]

Summary

This PR preserves the left operand layout for Add results when the output shape is unchanged and the operand is non-overlapping and dense.

It uses the public empty_like API with MemoryFormat::Preserve for that case and retains contiguous allocation for broadcasting, expanded inputs, and gapped layouts.

It extends the CPU-oracle test helper to copy raw noncontiguous result storage, converts the transposed-output strict expected failure into a passing conformance case, and adds an expanded-stride regression.

This advances the Add semantics work in #13.

Validation

• Ruff and clang-format 19.1.4 checks passed.
• PyTorch 2.12 and 2.13 CPU behavior probes agreed on the output layout policy.
• The complete installed-wheel CPU fallback suite passed with 143 tests passed, 32 skipped, and 2 expected failures.
• The NVIDIA wheel built, installed, and imported successfully. The transposed result reached value comparison with matching shape, stride, contiguity, dtype, layout, and storage offset metadata.
• The separate NVIDIA Add value instability is tracked in Stabilize NVIDIA Add validation #33 and is not treated as a layout regression in this PR.