API: Add Dictionary-encoded Extension Type

[TomAugspurger]

Currently, Categorical serves two main purposes

1. A type for expressing data from a fixed set of categories
2. A memory efficient storage format for low-cardinality objects

This proposal is to add a new extension type (let's call it DictEncodedArray
for now) for the second use case. The storage format would be the same as
Categorical: an Index of the unique "keys" (categories) and an array of codes.
Much of the implementation would be shared. But they would have different
semantics on operations

• concat (union by default)
• groupby (unobserved categories would be dropped by default)
• value_counts (unobserved categories would be dropped by default)

This is most useful for strings, but could even be useful for storing a large
array of 64-bit precision items (store the 64-bit items once, then use an int16
or int32 array for the codes).

[toobaz]

From #20583 , @jankatins :

If you have a defined order, where would you put the new categorical into the orders? E.g. good - middle - bad, where would you put 'extreme' -> if you care about order, then you also care about not adding new stuff

Sure: an exception can be when the new categories include the old ones, but that's a very specific case, and I agree the ability to add categories is less relevant for ordered categories (by "less relevant" I mean that it could for instance loose the ordering, or append the new ones according to their own order).

[toobaz]

This is most useful for strings, but could even be useful for storing a large
array of 64-bit precision items (store the 64-bit items once, then use an int16
or int32 array for the codes).

... or even arbitrary Python objects (for which you gain not only RAM, but possibly CPU when e.g. comparing)

Did I understand correctly that for the case of strings we could expect some significant advantage in storing them as a unique string and map (internally) categories to slices of it, rather than directly map categories to individual strings?

[TomAugspurger]

Not happening for 1.0