Investigate using 24 or 16 bit depth for training models for speed (instead of 32)

[GuitarML]

Currently, 32 bit floating point audio is used for training. There may be a speed benefit from switching to 24 or 16 bit without significant loss of quality. Investigate for optimal speed vs sound quality. If there is a speed improvement, consider making the bit depth a selectable option when training models.