Add lru_cache to _calculate_amat_eigenvecs to speed up pswf_operator

[tyler-a-cox]

I added lru_cache to _calculate_amat_eigenvecs to speed up pswf_operator in the case where it's called repeatedly for similar input values. One of the main use cases for pswf_operator is hera_cal.nucal.compute_spatial_filters which computes PSWF basis functions for filtering baselines with the same orientation. Sets of baselines within the same orientation use the same eigenvectors and eigenvalues to compute a set of PSWF basis functions, but they're never explicitly cached even though they're the most expensive part of the computation. I set the maxsize fairly low, since the pswf_operator function is called repeatedly by hera_cal.nucal.compute_spatial_filters for baselines which reuse the output of _calculate_amat_eigenvecs. I also kept the maxsize low because the returned matrix size of this function can get up to nfreqs x nfreqs