ENH: consider providing only-interpolate versions

[fzimmermann89]

Hi,

I was wondering if it might be feasible to provide functions for cuda and cpu that only do the interpolations for type 1 and type 2 nuffts. This would allow us to combine these with torch fftn calls to obtain a full nufft -- without having to worry about storing plans anywhere and abiding by torch's multiprocessing settings.
Furthermore, having these is also useful for density estimation in MRI.
Currently, if I am not mistaken, this is only implemented for cufinufft.

I am opening this issue here first, as it is somewhat related to #56 -- but maybe it would be better to open it in the main repository?

Cheers
Felix

[WardBrian]

Currently, if I am not mistaken, this is only implemented for cufinufft.

If so, we would need support in the upstream finufft project before we could really think about implementing this request here