[C++][Python] Optimize aggregate functions to work with batches.

Most compute functions work with the dataset api and don't load columns. But aggregate functions which are associative could also work: `min`, `max`, `any`, `all`, `sum`, `product`. Even `unique` and `value_counts`.

A couple of implementation ideas:

• expand the dataset api to support expressions which return scalars
• add a `BatchedArray` type which is like a `ChunkedArray` but with lazy loading