Ack htlc settlement commands after writing state

[pm47]

There are two issues:

1. because we forward commands before writing to disk in
   PendingRelayDb.safeSend in order to reduce latency, there is a race
   where the channel can process and acknowledge the command before the
   db write. As a result, the command stays in the pending relay db and
   will be cleaned up by the post-restart-htlc-cleaner at next restart.
2. in the general case, the channel acknowledges commands before it
   writes its state to disk, which opens a window for losing the command
   if we stop eclair at that exact time.

In order to fix 2., we introduce a new acking transition method, which
will be called after storing. This method adds a delay before actually
acknowledging the commands, which should be more than enough to solve 1.

I'm not sure we need that additional delay, because now that we
acknowledge the commands after storing the state, the channel should
lose the race most of the time.

[t-bast]

LGTM, I think the additional delay doesn't hurt, let's keep it.