Fix provenance of kwargs

[fritzo]

This extends ProvenanceTensor to handle kwargs, which is needed in PyTorch 1.10 because torch.nn.functional.linear's bias input is passed as a named kwarg rather than a positional arg. To clarify the code I've also refactored to use singledispatch with two new functions track_provenance() and extract_provenance().

Tested

• refactoring is covered by existing tests (test_inspect.py, test_autoguide.py, test_gaussian.py)
• tested locally with PyTorch 1.10

cc @ordabayevy

[fritzo]

Thanks for reviewing @eb8680!