No Metal Support

[icakinser]

Doesnt run on Apple Silcone or the GPU.

[syun88]

@icakinser I already tried, and it works on Mac.

Article (Japanese): [Link to Qiita article](https://qiita.com/syun88/items/20cf758344825e9bea5e)

MPS mode does not work, but CPU mode works fine. I tried switching from MPS to CPU, but the generated images were all black, so I changed it back to CPU mode.

[GitHub Pull Request Link](#4)

Key Changes:

Dependencies:

• Replaced decord with eva-decord for macOS/py3.10 compatibility, while keeping av==12.3.0.

Dataset:

• Lazily initialized the multiprocessing Manager to avoid crashes during macOS spawn.
• Added lock guards for cases when the Manager is unavailable.

Inference:

• Cleared caches on CUDA or MPS when available.
• Moved tensors to the current device instead of hard-coding .cuda().

Training:

• Set the gloo backend when CUDA is not available.
• Guarded DistributedDataParallel for CPU-only runs.
• Only set the CUDA device when CUDA is available.

Transformer:

• Guarded all_reduce calls when torch.distributed isn’t initialized (to handle single-process, MPS, or CPU cases).

Scripts:

• Increased the default image sampling batch size to 16 in test_sample_image.sh.

Notes:

• sample.py remains unchanged from upstream (still selects CUDA if available, otherwise defaults to CPU). No forced CPU/MPS avoidance is included in this PR.
• The goal is to provide a smoother setup experience on macOS and ensure safer execution in non-CUDA environments.