Workaround for performance bug in PyTorch with subclassed tensors

[warner-benjamin]

Resolves #3682 by adding CastToTensor callback as a workaround for performance bug in PyTorch with subclassed tensors.

Fixes 4cff258 by testing if learn.xb & learn.yb are tuples and applying the cast to Tensor if they are. (Unless I made a mistake in testing, b[:i] and b[i:] are tuples).

Unlike 4cff258, this PR adds the workaround as a callback so callbacks which use the input tensor type still can before it is casted to Tensor for the training performance increase.

It also allows turning off the Tensor casting by removing the callback should it ruin a workflow. Although anyone who does this is encouraged to reimplement their own custom callback which casts to Tensor to get the free training performance increase.

Currently CastToTensor.order is right before MixedPrecision.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[warner-benjamin]

I should have fixed the sync error, but if a nbdev _all_ isn't on one line, it won't be added to the module's __all__.

[jph00]

Much better - thanks!