Expose randomized SVD for CUDA

[kshyatt]

We currently aren't exposing this for svd_trunc but we should in order to allow users to have best performance

[kshyatt]

Circling back to this. Currently if you try to use CUSOLVER_Randomized as the algorithm for svd_trunc (e.g. wrap it in a TruncatedAlgorithm), things fail because svd_compact! doesn't work with CUSOLVER_Randomized. Fair enough. But a bigger problem is that for TensorMaps with non-trivial block structure, it's not so obvious how to do this in the most efficient way, given that we want to "allocate" the truncation among the sectors. Would be good to discuss/think a little bit about how to expose this because the performance benefits are pretty compelling.

[lkdvos]

See also QuantumKitHub/MatrixAlgebraKit.jl#225