csr matrix multiply in sum layer

[jjcmoon]

• sum layer can be formulated as sparse matrix multiply.
• gives small (10% - 20%) speedup.
• sparse matrix multiply doesn't work on MPS in torch, so there the old scatter reduce is used.

[VincentDerk]

The documentation and import indicates it is experimental code, would be good to indicate this when we add a semiring that uses _plain_sum_layer.

Do I understand correctly that right now there is no such semiring in get_semiring, so the sparse code is not used yet?

Looks fine to me.