Matrix batches

[tgymnich]

This PR would enable lu factorization in batches. However the MPSKernel only ever runs once for LU decomposition, trying the same for MatMul works just fine.

[tgymnich]

Would be nice to direct something like:
eachslice(X; dims=3) .* eachslice(Y; dims=3) or @einsum C[i,j,b] := A[i,k,b] * B[k,j,b] to the batched MPS matmul kernel. Does not seem like any GPU backend is doing so currently...

[maleadt]

eachslice(X; dims=3) .* eachslice(Y; dims=3) or @einsum C[i,j,b] := A[i,k,b] * B[k,j,b] to the batch MPS matmul kernel

Yeah, I'm surprised we haven't come up with anything convenient yet. There's https://github.com/FluxML/NNlib.jl/blob/1b30040fabadd41efa0d9dde5841b90f9f85cf2d/src/batched/batchedmul.jl#L4-L54, but that would require an NNlib.jl dependency.