`sparse_mm` should use `sparse.mm` by default for unbatched CSR matrices

[theo-barfoot]

Following the benchmark results which states:

• For CSR backward, torch spmm outperforms tsgu spmm in both time and peak memory; tsgu spmm incurs extra memory due to internal CSR→COO conversion for gradient computation.
• tsgu spmm supports CSR gradients contrary to current public PyTorch documentation statements.

And, the confirmation of the PyTorch documentation not accurately reflecting this, as raised in pytorch/pytorch#172550

The function sparse_mm should be updated to use the PyTorch function for unbatched CSR matrices.

[tvercaut]

To ease benchmarking in the future, we could consider adding an option to sparse_mm to manualy choose the backend (say between auto, torch, and tsgu). The risk is code bloat though...