Revisit CuPy FFT single precision upcasting

[garrettwrong]

In #1140 ASPIRE's CuPy FFT wrapper was made functional and covered by our test suite. During testing I found there were some numerical and determinism differences that made it difficult to generically use CuPy's FFT generically in place of Scipy or PyFFTW on the host. These were resolved by using doubles for all CuPy FFT calls inside our wrapper.

I believe if areas of the code sensitive to singles are addressed directly the upcasting could be removed in the future.

         I removed the comment.  We can hopefully revisit this after addressing areas of the code that are not very robust in singles (single precision filters for example).

Originally posted by @garrettwrong in #1140 (comment)