finufft/cufinufft interface rationalization (Python)

[janden]

Similar to #255, we need to align the interfaces for finufft and cufinufft in Python.

init

Plan.__init__(nufft_type, n_modes_or_dim, n_trans=1, eps=1e-06, isign=None, dtype='complex128', **kwargs)
cufinufft.__init__(nufft_type, modes, n_trans=1, eps=1e-06, isign=None, dtype=numpy.float32, **kwargs)

These are quite similar, but we'll need to figure out what our convention is for dtypes (should it be the real or complex or should we accept both). Here I think it makes sense for cufinufft to default to single precision even though finufft defaults to double precision. This should be more in line with what one would expect for CPU/GPU codes, but I'm open to suggestions here.

setpts

Plan.setpts(x=None, y=None, z=None, s=None, t=None, u=None)
cufinufft.set_pts(kx, ky=None, kz=None)

Obviously the name needs to change. (set_pts to setpts) but perhaps also the argument names. Should think about extending the interface to handle type 3 even if it's not implemented yet.

execute

Plan.execute(data, out=None)
cufinufft.execute(c, fk)

These behave quite differently in that data is always input for finufft, but for cufinufft either c or fk can be input (the other is output) depending on the type. Ideally, we should also allow the user to return an array.

Finally, we should consider renaming the cufinufft class to Plan similar to the finufft case.

[janden]

Closed by #290.