speed up _key2inds

[adampbeardsley]

Right now the limiting factor in accessing the data quickly is looking up the indices corresponding to a given key. This would be substantially reduced if we can store the key -> index mapping, rather than computing it every time it's needing.

[bhazelton]

Should be cached as it is called, not on read.

Also might be faster to assign everything once rather than searching through the baseline array for each key.

[mkolopanis]

We might be able to use functools.lru_cache to generate a cache of key to ind values. then we can have a function that just clears the cache of keys whenever the state of the object changes (add, read, select, etc)

[mkolopanis]

the cache I mentioned before will not work for us because the UVData object itself is unhashable.

[mkolopanis]

I have tried this successfully on a branch, but it would just add another dependency. https://pypi.org/project/memoization/