[Fix] Add error message for exceptions caused by large gradients

[yoshikisd]

In issue #33, an exception was raised when an SGD optimizer (coupled with a high learning rate) was used to perform reconstructions on a fancy_ptycho.py example. However, the error message does not clearly describe that the root cause of the problem may likely be related to excessively large gradients as pointed out in #33 (comment).

This PR aims to improve the readability of this error message.

[gnzng]

Another common cause for nan values in the loss are related to the chosen loss functions and their normalization (divide by 0, neg log). So maybe it makes sense to raise that error more upstream when a nan loss is returned by the loss function checked by t.isnan(), instead of try.

[allevitan]

I think this is too general of a catch. Let me propose something else when I get a moment