Fix type usage in opnorm for AbstractGPUSparseMatrixCSC

[ChrisRackauckas]

It requires a type not a value.

Found in https://github.com/SciML/OrdinaryDiffEq.jl/actions/runs/28556241206/job/84664223984?pr=3810#step:11:516

[ChrisRackauckas]

The other possible fix is to add a dispatch to csr_type(A::cuSPARSE.CuSparseMatrixCSC) = csr_type((typeof(A)), not sure which is preferred.

[kshyatt]

Fails look related now

[christiangnrd]

@kshyatt This function is defined according to the definition in the GPUArrays readme in CUDA.jl (takes in the array type), but not JLArrays (takes in the array), but it's used as defined in JLArrays in mapreduce

GPUArrays.jl/src/host/sparse.jl

Lines 1009 to 1012 in b94388f

	if A isa AbstractGPUSparseMatrixCSR && dims == 1
	A = csc_type(A)(A)
	elseif A isa AbstractGPUSparseMatrixCSC && dims == 2
	A = csr_type(A)(A)

I suggest we fix this in JLArrays and define csr_type(A<:AbstractGPUSparseMatrix) = csr_type(typeof(A)) (and the csc_type equivalent

[kshyatt]

Sounds good, should we do this in this PR?

[christiangnrd]

Sounds good, should we do this in this PR?

Yeah I'll push