Add Apple M-series (MPS) device support to Python scripts

[brianholland]

Summary

Most standalone Python scripts (.py) only detect CUDA or fall back to CPU, ignoring Apple Silicon's MPS backend. This means users with M1/M2/M3/M4 Macs always fall back to CPU, missing significant GPU acceleration.

Problem

~29 Python files use this pattern:

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

This should be updated to also detect MPS:

if torch.cuda.is_available():
    device = torch.device("cuda")
elif torch.backends.mps.is_available():
    device = torch.device("mps")
else:
    device = torch.device("cpu")

Additional considerations

• pin_memory=True in DataLoaders should be conditional on CUDA availability
• torch.compile should be guarded (limited MPS support)
• DDP scripts are inherently CUDA-only and should be left as-is

Scope

• All standalone .py training/inference scripts in ch04-ch07, appendix, and pkg/tests
• Excludes DDP-specific scripts (appendix-A DDP scripts, ch05/10 multi-GPU)
• Excludes notebooks (many already have MPS support)

🤖 Generated with Claude Code

[rasbt]

If someone reads this, this omission of mps was on purpose (see #625 #644 #442 #846)

Perhaps worthwhile revisiting some time in the future though.