Add support for multi-key Top-K in `BlockTopK`

The goal here is to support multi-key top-k for small segments (as described in #8369). Supporting this through decomposer in `BlockTopK` is just one of the options. #8369 will inform the approach to take here.

From #8369, we came to the following conclusion:
> Regarding BlockTopK we decided to opt for an interface that allows consecutive refinement via multiple applications of BlockTopK which gives more flexibility to the user to not keep all key-fragments in on-chip memory (vs. a Decomposer-based interface). To make this possible the interface needs to accept and return an object that keeps track of the status of each item/index (selected vs candidate/ties).

_Depends on:_
- #8369