Revise API expectations to enforce Tensor inputs to distributions

[matthewfeickert]

In PR #1351 we needed to add self.astensor calls back into the PyTorch backend, which we worked hard to remove in PR #582. This is the result of the public API expecting that you can pass floats and not tensors to the distribution methods and still have things work.

We should change this expectation in the API and docs to make it clear that all inputs must be tensors, and then remove the self.astensor calls.

Originally posted by @matthewfeickert in #1351 (comment)

[matthewfeickert]

@kratsg rightly points out in PR #1399 that

We need to check that the fit times for tflow don't change.

given the self.astensor calls that were added to the TensorFlow backend there as well.