Batched diagonalization on CUDA GPUs

[kbarros]

CuSolve provides a function to perform batched diagonalization of Hermitian matrices:
https://docs.nvidia.com/cuda/cusolver/index.html#cusolverdn-t-syevj

Performance benefits may depend a lot on matrix size, etc: https://discourse.julialang.org/t/eigenvalues-for-lots-of-small-matrices-gpu-batched-vs-cpu-eigen/50792

We could consider using this for accelerating LSWT. Two concerns:

1. What we actually need is a batched generalized eigenvalue decomposition. That is, we need sygvj and not syevj, as documented here. However, the existing batched eigensolvers do not seem to support the generalized case? We could implement a generalized eigensolver ourselves using batched Cholesky decomposition, which is supported by CUDA.
2. Given that many LSWT calculations will typically working on small matrix sizes, especially in dipole mode, the diagonalization subroutine itself may not be the dominant cost. To make this beneficial, we would probably need to move a lot of the calculation onto the GPU (e.g., the matrix-builds for each q).