Better workarounds for MPS bug

[sdatkinson]

• Exclude PyTorch 2.5.0 and 2.5.1 from requirements. (TBD 2.5.2).
• Keep the workaround and print a longer error message if the chunking fallback is used so that folks have more info to relay. I want to hear if this is still a problem in future versions of PyTorch.

[sdatkinson]

On second thought, adding constraints to NAM's dependencies may cause it to be harder to integrate with other code. That's always a pain, so I'm not going to impose extra restrictions on its requirements since the workaround works (but at a performance hit--and usually only in validation).

If folks want to, they can dodge the versions of PyTorch that are problematic with this in conda YAML:

  - pytorch<2.5.0
  - pytorch>2.5.1. # And possibly more--we'll see.

or this in pip:

torch!=2.5.0,!=2.5.1