Allow for either 'nvrtc' or 'nvcc' backends

[wmwv]

Add the ability to support either nvrtc or nvcc cuda backends.

1. Minimal: Set variable to replace all of the currently string literal "nvcc". This could still be hardcoded, but would be easier to update in the code.
2. Better: Have a configuration parameter that can be passed to sfft to choose which backend.
3. Stretch: Set up config system handling following snappl/photometry examples. This is probably too much effort for something that doesn't matter and would start to customize this fork unecessarily.

Motivation: The 'nvrtc' backend doesn't require that the CUDA version support the available gcc compiler. The 'nvcc' does; and furthermore, GPUs break containerization abstraction because the CUDA version has to match the driver in the host OS running the hardware. So as @rknop and @wmwv have been testing out different containers (on NERSC), it is (a) helpful to have this backend flexibility and (b) at least changing to 'nvrtc' is necessary to support desired combinations right now on NERSC.

[wmwv]

Relevant files are:
sfft/BSplineSFFT.py
sfft/sfftcore/SFFTConfigure.py
utils/ResampKits.py

I'm defining CUDA_COMPILER to go along with the BACKEND_4SUBTRACT and -CUDA_DEVICE_4SUBTRACT arguments that are passed around.

[wmwv]

Implemented 1+2.