refactor(python): migrate torch.jit.script to torch.compile

[wjones127]

Closes #5762

Summary

• Replace @torch.jit.script with @torch.compile in distance.py (6 functions)
• Add torch>=2.0 version constraint
• Add filterwarning for PyTorch inductor's use of deprecated load_module()

Test plan

• pytest python/tests/torch_tests/test_distance.py passes

🤖 Generated with Claude Code

[github-actions]

Code Review

Summary

This PR migrates from the deprecated torch.jit.script to torch.compile for PyTorch 2.x compatibility.

P0/P1 Issues

P1: Behavioral differences between torch.jit.script and torch.compile

The migration from @torch.jit.script to @torch.compile changes compilation semantics:

• torch.jit.script compiles eagerly at decoration time
• torch.compile compiles lazily on first call and uses the inductor backend

This could cause:

1. First-call latency: Initial invocation will trigger JIT compilation, which may be significant for these distance functions
2. Dynamic shape handling: torch.compile by default recompiles on shape changes. For functions like _l2_distance and _cosine_distance that are called in loops with different split_size values, this could trigger multiple recompilations

Consider adding torch.compile(dynamic=True) if shape variation is expected, or documenting the first-call compilation overhead.

Minor Observations

• The existing tests cover the main functions but may not stress-test the compilation behavior with varying input shapes
• The typo fix in the comment ("Python 2.12" → "Python 3.12") is a nice catch

Verdict

The change is reasonable for PyTorch 2.x compatibility, but please verify that the dynamic batching logic in cosine_distance and l2_distance (which vary split_size) doesn't cause excessive recompilation overhead.