Upload model checkpoints to S3

Right now they're only saved to local disk. Would be nice to have the option to persist them remotely. Although... maybe to disk synchronously and then to S3 in a background thread?

See https://github.com/eihli/mugato/blob/1deec8096c96f94eab5017161fa97d2e40f80a4f/mugato/train.py#L322