Optimized decompression for specific matrix types

[gdalle]

At the moment, our only optimized decompression is for SparseMatrixCSC in :direct mode: we store a vector of compressed_indices such that nonzeros(A) = vec(B)[compressed_indices].
We can probably find a similar optimization for :substitution mode.

What do we want to do for other matrix types, like:

• the ones from LinearAlgebra: Bidiagonal, Tridiagonal, Symmetric, etc.
• BandedMatrices.jl, BlockBandedMatrices.jl

It would be rather tiring to find optimal decompression methods for each of these. My proposal (as a first step) would be to always have a SparseMatrixCSC buffer into which we decompress, and then copy the A_buffer::SparseMatrixCSC into A::SomeWeirdMatrix.
Essentially, it's easier to implement fast copy from SparseMatrixCSC than fast decompression.

Related:

• Optimized coloring for specific matrix types #65
• Provide eltype of decompression target A #44

[gdalle]

Conclusion of our discussion:

• Keep one decompression for SparseMatrixCSC and one decompression for AbstractMatrix to remain alloc-free in both cases

[gdalle]

See also:

• Performance Regression from SparseDiffTools: Specialized sparse matrix decompression for structured matrices JuliaDiff/DifferentiationInterface.jl#780