Series index: don't rely so heavily on NOSQL to dedupe writes.

[tomwilkie]

The series store will write the index entries (metric:label -> series) for every chunk, and rely on the underlying store to dedupe these. We could keep a cache of the most recently written entries and not bother writing ones we already know exist, removing some write load on the database.

For an ingester with ~1m series, an average chunk size of 6hs, and 10 labels per chunks, we're writing ~500 entries / sec. A cache of 10m entries would potentially reduce this to ~entries / sec.

[bboreham]

Somewhat related to #607