ci(pt_expt): cover torch.compile on a supported GPU release #5935

Problem

The repository-wide GPU CI environment currently pins PyTorch 2.10, while the supported torch.compile releases are PyTorch 2.11 and 2.12, with PyTorch 2.13 support added by #5910. Consequently, pt_expt compile-only tests cannot run on GPU CI under a supported PyTorch release.

This supported-version GPU coverage gap predates #5913. That PR only makes the existing contract explicit by using the shared compile-version gate, so the PyTorch 2.10 CUDA job now skips compile-only tests instead of exercising an unsupported configuration.

The compile path has been manually validated on CUDA with PyTorch 2.11, 2.12, and 2.13. The repository-wide GPU pin cannot currently be raised because of unrelated CI compatibility issues, and changing that pin is outside the scope of #5913.

Desired coverage

• Run the focused pt_expt compile tests on a GPU with at least one supported PyTorch release.
• Exercise CUDA-specific compile safeguards such as forced int64 indexing and the constrained Triton tiling configuration.
• Preserve the existing PyTorch 2.10 eager CUDA coverage until the unrelated CI blockers allow the global GPU pin to move.

Related

• perf(pt_expt): use scalable graph builders during training #5913
• fix(pt): calibrate DPA4 bias and support PyTorch 2.13 compile #5910