Don't call check_residual! in batch

[klamike]

@frapac @amontoison What do you think? It saves a bunch of time since we avoid the KKT mul and norm calls. Note this is merging into MadNLP#81

Maybe we can replace it with a cheaper check (like just finite/non-nan solution?) instead.

For reference,

function check_residual!(d::MadNLP.UnreducedKKTVector{T}, solver, p) where T
    opt = solver.opt

    # Check residual
    w = solver._w1
    copyto!(MadNLP.full(w), MadNLP.full(p))
    mul!(w, solver.kkt, d, -one(T), one(T))
    norm_w = norm(MadNLP.full(w), Inf)
    norm_p = norm(MadNLP.full(p), Inf)

    residual_ratio = norm_w / max(one(T), norm_p)
    MadNLP.@debug(
        solver.logger,
        @sprintf("Residual after linear solve: %6.2e", residual_ratio),
    )
    if isnan(residual_ratio) || (opt.check_residual && (residual_ratio > opt.tol_linear_solve))
        throw(MadNLP.SolveException)
    end
    return d
end

[amontoison]

@klamike I think we have batch spmv and spmm in CUSPARSE. I interfaced it or reviewed it but it is not documented.
For the norm calls, we can have asynchronous ones, NVIDIA added it in CUDA 12.x.
The current issue is that all operations are blocking so we multiply the elapsed time by the number of batch.

I can work on an operator for batch mul!, I understand quite well the internal of CUSPARSE routines.

References:

• Switch CUBLAS to device-side pointer mode JuliaGPU/CUDA.jl#2616
• CuSparseArrayCSR (N dim array) with batched matmatmul (bmm) JuliaGPU/CUDA.jl#1944

PS: I should have used CuSparseArrayCSR in CUDSS...
I didn't remember that we added this type.

[klamike]

Let's close this then, and keep the check_residual! open on the tracker.