storage: Asynchronous snapshots

[bdarnell]

Now that Replica.Snapshot is safely isolated from the rest of the Replica (#6187), we can use etcd/raft's asynchronous snapshot feature. This will avoid blocking the processRaft goroutine during the expensive snapshot generation.

The first time Snapshot() is called, it should start a goroutine and immediately return raft.ErrSnapshotTemporarilyUnavailable. Raft will occasionally poll for results by calling Snapshot; we can return the result when we have it; until then we return ErrSnapshotTemporarilyUnavailable again (without starting new goroutines).

A few subtle points:

• In rare cases raft may decide it doesn't need the snapshot, so we should be sure to discard snapshots that go unused for too long.
• When expanding the replication factor (e.g from 1 to 3 or 3 to 5) we may be able to reuse the same snapshot twice, although It's probably not worth optimizing for this case.
• After starting the goroutine, it might be best to wait with a short timeout so that we can get the snapshot in a single attempt when it's small. This will be especially important to keep the tests fast.