GPU-friendly preconditioning

[gdalle]

Currently, the preconditioning step requires a "power sum" over columns or rows of a sparse matrix. It is very likely that this operation is not implemented efficiently for all the matrix types we care about (even though we have a good SparseMatrixCSC version).
If we replace this operation with a map (raising to the correct power) followed by a multiplication with a vector of ones, we might get something that most GPU sparse matrices support out of the box

[klamike]

some related work: https://github.com/klamike/MathOptScaling.jl

don't depend on it though, it won't be getting any more updates since I plan to merge it into https://github.com/klamike/BatchQuadraticModels.jl (perhaps we should consider adopting that, though for CoolPDLP since we are so focused on code quality/clean abstractions, maybe best to roll our own)