Standardized trainer: Leave cab-fitting loss on all the time?

[sdatkinson]

Instead of having cab modeling/fit_cab be an option, perhaps it should just be on all the time.

Pros:

• Simpler. "It just works".
• You can always use bin/train/main.py to customize and not use it if you want.
• Can roll it back if something else goes wrong.

Cons:

• Harder to tell if the model includes a cab.
• Slower training, esp on macOS that doesn't support the MRSTFT ops (though perhaps this could be worked around better than today). One quote I got was that it's a 4-5x difference.
• Slightly higher ESRs, usually.

Seeking other insights about this--what other considerations are there?

[sdatkinson]

This would severely impact training on macOS with Apple Silicon due to the lack of support for MRSTFT loss on the accelerator.

I'm not going to ship this until I can improve the situation there. However, once that's fixed, I'll ship this (and users who want it back can use an extension)