Fix default_device to correctly detect + use mps (Apple Silicon)

[wolever]

This fixes two minor bugs preventing fastai.torch_core.default_device from detecting mps (Apple Silicon) GPUs.

1. The torch.backends.mps.is_available attribute does not exist in PyTorch 1.13.1. If this attribute is unavailable, this patch uses torch.has_mps instead.
2. A minor logic bug in default_device prevented _has_mps from being called.

Additionally, this fixes #3785 by removing the duplicate default_device function.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[jph00]

Many thanks.

[ypauchard]

@wolever the new _has_mps() function now in fastai 2.7.11 causes an error on Intel Macs with OS < 12.3:

RuntimeError: The MPS backend is supported on MacOS 12.3+.Current OS version can be queried using `sw_vers`

I think it is because torch.backends.mps.is_available() returns False (the correct value), but torch.has_mps is True. In this case _has_mps() returns True and the backend is set to mps. Using mps will cause the error above.

I suggest changing _has_mps() to return the value of torch.backends.mps.is_available() if it exists and only default to torch.has_mps if this function is not available