fix: --device autodetect the available backend

[rizzler13]

Closes #396

Changes added:
Adds a small _get_default_device() helper to each of the 3 CLI entry points that auto-detects the best available backend:

• torch.cuda.is_available() → "cuda"
• torch.backends.mps.is_available() → "mps" (Apple Silicon via Metal)
• fallback → "cpu"

Files changed:

• moshi/moshi/server.py — added _get_default_device(), updated --device default, guarded torch.cuda.synchronize()
• moshi/moshi/run_inference.py — added _get_default_device(), updated --device default
• moshi/moshi/run_tts.py — added _get_default_device(), updated --device default

Explicit --device cuda still works exactly as before. The only behavioral change is the default when --device is omitted.

Note: seed_all() is duplicated identically in server.py and run_inference.py. It's already correctly guarded (if torch.cuda.is_available()) so it's not part of the crash path, I left it untouched to keep this PR focused. Happy to consolidate it in a follow up if that'd be useful.

[nouraellm]

LGTM! The only thing I'd verify is that the existing default dtype (torch.bfloat16) behaves correctly when the auto-selected device is mps or cpu. Otherwise I don't see any issues with the change.

[rizzler13]

Just double checked this on my machine to be sure,
both cpu and mps work fine with the default configs (bfloat16/float16), and nn.Linear runs without any issues.