Need to select TPL algorithms

[jczhang07]

More specifically, if TPL CUDA is enabled, I want to try various cusparse_spgemm algorithms, including

    CUSPARSE_SPGEMM_DEFAULT                 = 0,
    CUSPARSE_SPGEMM_CSR_ALG_DETERMINITIC    = 1,
    CUSPARSE_SPGEMM_CSR_ALG_NONDETERMINITIC = 2,
    CUSPARSE_SPGEMM_ALG1                    = 3,
    CUSPARSE_SPGEMM_ALG2                    = 4,
    CUSPARSE_SPGEMM_ALG3                    = 5

but KK's SPGEMMAlgorithm does not support them (except DEFAULT).

[lucbv]

Starting to look into this. Might want to change the handle and the no_reuse version of the TPL implementation layer.
I don't think these are callable from the spgemm implementation that allows reuse.
It looks like we will need to check for CUDA_VERSION >= 12001.
Finally, we will need a couple of implementation for the regular variants and the mem variants.