Fix bug in setting training data for ExactGP if training inputs are None

[neighthan]

It appears valid for a GP to have its training inputs/labels set to None. However, if you call set_train_data when the current training inputs are None, you'll get an error at for input, t_input in zip(inputs, self.train_inputs) (because None is not iterable and thus can't be zipped). This PR moves the check for if strict up so that, if strict == False, the entire checking section is skipped and it doesn't matter whether the training data was None or not.

When strict == True, the error messages have been improved a bit. Now they'll say both what the expected attribute (e.g. shape) was and what shape was found. To get a more informative error, I also ensured that if the training inputs/targets are None, the shape, etc. will be viewed as None so you'll still get the expected RuntimeError (instead of an error at zip that is harder to diagnose).

This passed flake8 fine, and I think the only errors I got in the tests were for CUDA out of memory.

[neighthan]

I also added a bit of documentation for the arguments of set_train_data (since I came to look at the source code to figure out what strict would do when I needed it). I didn't specify shapes for inputs and targets - should these be the same as in get_fantasy_model? E.g.

"""
- :attr:`inputs` (Tensor `m x d` or `b x m x d`)
- :attr:`targets` (Tensor `m` or `b x m`)
"""

[jacobrgardner]

This looks great! Thanks @neighthan