Update GPU implementation to take advantage of multiprocessing

[devincody]

The GPU implementation is currently bottlenecked by the time that it takes to generate orbits which are then solved on the GPU. There are two solutions that could make the GPU implementation worth using.

1. Generate the orbits on the GPU. This would require moving a large portion of the orbitize code from python to C/CUDA which would make the code more difficult to maintain.
2. Run the GPU implementation out of a single thread and use the rest of the threads in the C solver mode. This was partially implemented in the pycuda_mp branch.

[sblunt]

Thanks Dev! Deleting the pycuda_mp branch for now, but we'll still be able to access it if/when we want to look back at what you tried in the futue.