Move cuSolver handle into CUDA runtime

Pretty straightforward cleanup---instead of making and deleting a handle every time chol() or svd() is called, just make it in the runtime.