REF/ENH potentially make nn the more underlying set of objects or figure out a canonical way of storing plans in pytorch functions

[eickenberg]

finufft is based on fftw, which computes ffts in two steps. First it makes a plan, then it executes the plan. When running many ffts it is important not to recompute the same plan for every FFT, since this will cause a significant slowdown.

This is more of a next-generation stretch goal at baseline, but if we discover that computing plans takes significant time in high dimensions, we should consider exposing objects that allow caching them.