Require that the segment has enough dirty documents to create a clean
chunk before recompressing during merge, there must be at least maxChunkSize.

This prevents wasteful recompression with small flushes (e.g. every
document): we ensure recompression achieves some "permanent" progress.

Expose maxDocsPerChunk as a parameter for Term vectors too, matching the
stored fields format. This allows for easy testing.

If segment N needs recompression, we have to flush any buffered docs
before bulk-copying segment N+1. Don't just increment numDirtyChunks,
also make sure numDirtyDocs is incremented, too.

This doesn't have a performance impact, and is unrelated to tooDirty()
improvements, but it is easier to reason about things with correct
statistics in the index.