A[c|t]_mul_B[!] specializations for SparseMatrixCSC-StridedVecOrMat, less generalized linear indexing and meta-fu

[Sacha0]

Towards fixing some generalized linear indexing deprecation warnings in #20040, this pull request rewrites the A[c|t]_mul_B[!] specializations for SparseMatrixCSC-StridedVecOrMat combinations, leveraging multiple dispatch and higher order functions to remove generalized linear indexing and meta-fu. cc @mbauman. Best!

[andreasnoack]

Could you explain the connection to generalized linear indexing and why it fixes the issue?

[Sacha0]

Could you explain the connection to generalized linear indexing and why it fixes the issue?

The existing implementation uses generalized linear indexing (when indexing B and C), whereas the implementation in this PR does not (by providing separate kernels for StridedVector B/C and StridedMatrix B/C). Removing these uses of generalized linear indexing fixes some depwarns in #20040, e.g. those appearing in the sparse/spqr tests. Does that clarify? Best!

[andreasnoack]

So the problem is indexing a vector with two indices? I thought that generalized linear indexing was about indexing an N dimensional array with n indices where 1 < n < N. The old behavior (which I though we called trailing singleton dimension) is pretty convenient in this case where it allows us to avoid duplication of the kernel.

[Sacha0]

So the problem is indexing a vector with two indices? I thought that generalized linear indexing was about indexing an N dimensional array with n indices where 1 < n < N. The old behavior (which I though we called trailing singleton dimension) is pretty convenient in this case where it allows us to avoid duplication of the kernel.

@andreasnoack, please see #20040. Best!