Refactor doc-id encoding for DiskBBQ to allow doc ids to be encoded in blocks

[benwtrent]

This uniformly encodes a bunch of doc-ids with the "worst" encoding. This way we store the encoding value just once, and can then encode the doc IDs uniformly and decode them in chunks at a time.

This is currently unused, but there are two things that I think that could benefit from this:

• doc id ranges being encoded for centroids (eager/lazy filtering)
• encoding doc ids along with vector blocks to prevent potentially loading large integer arrays for overly large clusters

[elasticsearchmachine]

Pinging @elastic/es-search-relevance (Team:Search Relevance)

[iverase]

This change makes sense to me.