Bfloat16 quantization

[mrariden]

CPSAM's transformer weights are unnecessarily too precise at 32bit, and 16 bit is sufficient for prediction. Switching to 16bit comes with multiple benefits, not the least of which is to free up RAM for loading more images for evaluation, reducing OOM issues.

This PR:

• Sets default to bfloat16, although this can be changed during CellposeModel object instantiation using the use_bfloat16 flag.
• Reduces model size from 1.2GB to 580MB
• Reduces runtime by ~20%
• Retains all segmentation accuracy

Testing:

• Notebook verification for API
• GUI tesing
• CLI testing