Reduce per-key metadata overheads

[vsreekanti]

There are a few places in which we could reduce metadata overheads:

• Instead of tracking key access frequencies for every single key in the database, we should instead use a heavy hitters sketch to track the most frequently accessed keys.
• We shouldn't explicitly cache and materialize the replication factor for a key whose rep factor is the default rep factor.
• Depending on how we use key sizes and what for, it might make sense to store a histogram of key sizes rather than every single size.