Need to compile Numba-CPU first to be able to compile Numba-GPU per #229

[ilhamv]

This behavior shows up per #229, which factored local arrays and objects out to a dedicated local.py file to avoid problematic circular imports for future development

[ilhamv]

Some finding that may be useful to debug:
If we start clean (no CPU/GPU cache at all), python input.py --mode=numba --target=gpu will fail (the error message seems to be the one for Numba-CPU), but the __ptxcache__ is still generated.
If we then do python input.py --mode=numba, it compiles and runs as usual.
If after that, we proceed with python input.py --mode=numba --target=gpu, which will use the existing __ptxcache__, it runs alright.

[ilhamv]

This is addressed per #241