Memoization scheme

[tlambert03]

As @fjug mentioned in #23 (comment)

It would be nice to cache/memoize expensive operations in general, such as generation of ground truth, PSFs, etc.

In some cases, where randomness is involved, this has to be done with care, and those functions should perhaps be unmemoizeable without providing a random seed. But for things like PSFs it's straightforward.

An additional complication is the numpy backend generalization: we don't want a numpy array returned if someone then switches to a Jax backend... so that needs to be part of the memo.

Since the objects can be large, global memory limits need to involved (this is particularly tricky if we're memoizing GPU-device objects... which probably
shouldn't be allowed)