cuSOLVER wrappers should not allocate a new handle every time

[ianmccul]

Currently in the GPU linear algebra wrappers that use cuSOLVER, the wrapper constructs a new cusolverDnHandle_t object every time. This is a slow operation, that should only be done once in each CPU thread that calls cuSOLVER.

Exactly how this is done is a design decision that needs to be addressed.