This is a follow up to
OAK-3349, which introduced tail compactions:
The deduplication caches currently only take the full generations into account and ignore the tail generations. Cache generations need to be a monotonically increasing, ordered sequence consisting of the full and tail part of the gc generation. See FileStoreBuilder.EvictingWriteCacheManager.evictOldGeneration. Optimally we find a way to decouple the segment generation from the cache generations as these are really separate concerns.