Check "Autograd" code generation custom operation edge case is covered by tests.

[pgmoka]

Create tests in https://github.com/pytorch/xla/blob/master/test/test_manual_xla_registration.py to verify “autograd” code generated operations are lowered as expected. In #8713 we saw how the interactions between “autograd” code generated operations and custom operations create an edge case where PyTorch might reach out to the "XLA" dispatch key when we didn't explicitly lower anything there.

We can manually resolve this issue as in the case for einsum, but since other operations do not seem to be as problematic, the solution from #8713 might be overkill. Instead, we should add test in https://github.com/pytorch/xla/blob/master/test/test_manual_xla_registration.py to make sure there is no shift from the PyTorch implementation