Introduce Abstract types for sparse arrays

[albertomercurio]

I find the AbstractGPUArray implementation very useful, which allows to have very generic methods for any GPU array, from CUDA.jl to Metal.jl and others.

Here, I try to implement such feature for sparse arrays. It is still a draft, but the basic method seems already to work. I can perform matrix-vector multiplication between a JLSparseMatrixCSC and a `JLVector.

What I would like to implement in this PR:

• Abstract types for all the sparse arrays
• Sparse Vector using JLArrays
• CSC sparse matrix using JLArrays
• CSR sparse matrix using JLArrays
• COO sparse matrix using JLArrays
• vector-vector dot product
• matrix-vector multiplication with CSC types
• matrix-vector multiplication with CSR types
• matrix-vector multiplication with COO types
• matrix-matrix multiplication with CSC types
• matrix-matrix multiplication with CSR types
• matrix-matrix multiplication with COO types
• Add tests

I think this is already enough for a single PR.

[maleadt]

Interesting!

FYI, CUDA.jl already has quite some native sparse operations implemented (including broadcast), so you may want to look there for inspiration, or to port additional functionality.

[albertomercurio]

Yes I'm already using some of them. I recently needed to implement sparse support for Metal.jl, so I decided to implement it very generally here.

I will for sure take inspiration from the CUDA sparse implementation.

[cortner]

Sparse Matrix x dense vector and Sparse Matrix x dense matrix (few columns) for just the CSC format would already cover many relevant use cases and imo would be worth merging without the ambitious complete set of features.

[kshyatt]

Is this still live given the merged sparse changes?

[albertomercurio]

Sorry for my late response. I didn't have time to continue it, and in the last period I decided to make a library that does it directly

https://github.com/albertomercurio/DeviceSparseArrays.jl

[kshyatt]

Can I close this PR, then?

[albertomercurio]

Yes