Unconditional Pebble DeleteRange on every Raft save stalls the cluster
A Raft log store that writes DeleteRange(last+1, MaxUint64) on every Save, to drop a possible conflicting tail, leaves a range tombstone even on plain appends. All of these tombstones end at the same key, so Pebble has to fragment thousands of nested ranges on each memtable flush, and that work grows quadratically. Under sustained load (~38 s at 400/s) every shard stalled, held elections and grew to about 2 GiB. The fix is to track the last entry index per store, loaded on open from the last entry key or the snapshot index if no entry follows it, and only write the range delete when the new entries end before that index. A good regression test does many small appends and fails if any single append takes too long: on the old code one append took 2.3 s.