storage,kv: consider reorganizing local keyspace keys to reduce w-amp

Within the range local keyspace, we have some keys that are unversioned (eg, RaftHardState, RaftAppliedState, RaftTruncatedState) but frequently written. With expiration-based leases, the lease key is also very frequently written. Today these keys are not stored continuously; replica raft logs, replicated shared locks, etc are interspersed. This interspersing increases write amplification.

Consider if we stored all the lease keys together in a continuous lease key prefix and imposed sstable boundary guards (cockroachdb/pebble#517) on either side. With expiration-based leases enabled, every memtable flush would typically write 1 lease key per replica to L0. Due to sstable boundary guards, this L0 file should be able to move-compacted to L5 and then compacted to L6 completely replacing the contents of the L6 sstable. This would reduce the write amp for the lease keys themselves to 3 (to WAL, to L0, to L6).

Other high-write, unversioned local keys might also benefit from a similar approach.

I'm uncertain the w-amp savings are enough to warrant the complexity of a migration.

Jira issue: CRDB-35891