Separating out Sparse matrix capability into separate repo

[ViralBShah]

As was done with Statistics.jl in #33399, it would be a good idea to do the same for sparse matrix capabilities. Are there any arguments against doing this?

I propose moving stdlib/SparseArrays and stdlib/SuiteSparse into external repos.

[ViralBShah]

I am not sure how to provide binary dependencies since those files go into deps. Perhaps should also wait for @staticfloat 's work on #35193 which will address this, before starting on this path.

[andreasnoack]

I think we'd have to deprecate some methods in LinearAlgebra before we can do this. E.g.

julia> similar(Bidiagonal(ones(4), ones(3), :L), (4,3))
4×3 SparseArrays.SparseMatrixCSC{Float64,Int64} with 0 stored entries

I also think there are some broadcasting combinations of special LinearAlgebra that produce a SparseMatrixCSC but I cannot reproduce an example on top of my head. Somebody should probably just try to remove SparseArrays and SuiteSparse and run the LinearAlgebra tests to see what fails.

[KristofferC]

Note that this issue seems to only be about the place where the code for the SparseArrays stdlib resides.

It doesn't change anything when it comes to what goes into the sysimage.

[andreasnoack]

It doesn't change anything when it comes to what goes into the sysimage.

Thanks for clarifying that. Indeed, I assumed that it would no longer go into the sysimage.

[ViralBShah]

Note that this issue seems to only be about the place where the code for the SparseArrays stdlib resides.

It doesn't change anything when it comes to what goes into the sysimage.

That's right.

[ViralBShah]

We also want to eventually move SparseArrays and SuiteSparse out of the system image, to pave the way for a variety of other sparse matrix improvements.

[ViralBShah]

Dup of #22698