CLI robustness: output path checked only after processing, errors go to stdout, GPU hard-coded to device 0

[joeljose]

Severity: Medium. On the MIT samples, a CPU run takes over an hour (see the performance issue), so a failure at the very end is expensive.

Problems

1. The output path is checked last. save_wav is the final step, so a directory that doesn't exist or can't be written raises FileNotFoundError/PermissionError as a traceback after all frames are processed, and the recovered audio is lost.
2. Errors go to stdout. Every Error: / Warning: message uses plain print(). Scripts and pipelines can't separate diagnostics from output. (The VRAM warning is the only one sent to stderr.)
3. GPU device 0 is hard-coded. torch.device('cuda'), get_device_name(0) and mem_get_info(0) are all fixed, and there's no --device flag, although the sibling projects have one.
4. estimate_vram ignores its nlevels argument. It's a fixed 15× multiplier, and it only warns, so an OOM still happens later.
5. No torch.no_grad(). The GPU forward pass runs without torch.inference_mode(). It's harmless today because the filters are buffers, but a future pytorch_wavelets could make them parameters and track gradients.
6. Timing. Pipeline timing starts before validation, and the CPU path imports dtcwt twice.

Suggested fix

• Resolve and check the output path at the start of main(): create the parent directory or fail, and do a test write with os.access.
• Wrap the final write so a failure saves to a fallback path such as ./sound_<timestamp>.wav rather than losing the result.
• Send all diagnostics to sys.stderr, or use logging. Add -q/--quiet.
• Add --device cpu|cuda[:N], used for model placement, get_device_name and mem_get_info. It also allows CPU-torch runs in CI.
• Use with torch.inference_mode(): around the forward and phase maths.
• Either use nlevels in the estimate or drop the parameter. Catching OOM once and halving the batch automatically would be friendlier than exiting.

Acceptance criteria

• An unwritable output path fails in under a second, before any frames are read.
• Errors go to stderr.
• --device is honoured.