Improper Appenderator.add() calls concurrent with persist

This issue is to track NPE reported in https://github.com/RoaringBitmap/RoaringBitmap/issues/358 but it doesn't seem that it may be caused by a bug in RoaringBitmap. Hence I think it's caused by a bug in Druid.

@jihoonson

> the read operation (persist operation in Druid) happens only when all write operations (indexing operation in Druid) are finished.

Could you demonstrate which synchronization mechanism or a happens-before edge or a casualty edge specifically ensure this? I cannot find it. It seems to happen somewhere in `SeekableStreamIndexTaskRunner.runInternal()` (related: https://github.com/apache/incubator-druid/issues/7360).

Start by reading SeekableStreamIndexTaskRunner.runInternal() and the related issue linked in the report, then trace how indexing writes overlap with Druid's persist operation. Done means either demonstrating the synchronization or happens-before edge that prevents concurrent Appenderator.add() and persist calls, or confirming and documenting the NPE fix path.

Written by the indexing model from the issue text.