We currently blanket disallow month or year intervals when passing operations down to cftime:

cftime disallows these intervals, presumably because it requires consistent-length time increments. But cftime is more intelligent about this than cf-units is, e.g. month intervals are allowed for 360_day calendars (Unidata/cftime#75).

We should make sure our own checks don't cover over details like the 360_day exception. The Pythonic choice would be to do no extra checks at all in cf-units, simply relying on cftime to raise the appropriate error; is there a good reason why we don't do this?