The floor raised for host so far (0 when none).
Never go faster than ms on host again, whichever engine asks. Only raises.
Wait for host's turn: after every earlier caller, and at least intervalMs
after the last request to it. Records the request either way, so an engine
with no interval still counts for one that has one.
The per-host pacing book: when each host was last asked, and a queue per host so that requests to one host are spaced out one after another even when several callers (engines, or concurrent requests of one engine) want it at once.