Sparse CSR: enable autograd for torch.sparse.addmm and torch.sparse.mm
This PR updates the derivative rule for `torch.sparse.addmm` to be
working with CSR sparse matrix. Notably `torch.sparse.sampled_addmm` is
used in the backward function.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/76591
Approved by: https://github.com/cpuhrsch
