Add 32-bit support in CUDA sparse solvers

[luisenp]

[maurimo]

Not possible at all in LUCuda (cusolverRfBatchResetValues only accepts double values), but should be completely trivial with BaSpaCho (you can even do the symbolic factorization, and decide later to factor either float or double numerical data).

[mhmukadam]

That's unfortunate about LUCuda, maybe we can just raise warning (before converting to double) if we aren't doing it already.