perf(gateway): optimize CommandRatelimiter size

[vilgotf]

• stores relative instead of absolute timestamps, going from 12 bytes to 2 bytes per item
• implicitly store the first relative timestamp in delay (if delay > now, one permit + pending.len() is acquired, else zero permits are acquired)
• always store waker (call Sleep::poll) in poll_pending
• fix theoretical issue from slice::partition_point returning an arbitrary element in the set of elements matching the predicate (we depend upon the first being returned)

TODO:

• Update doc comments
• Explain algorithm more in code comments, the logic is complex
• (maybe) add more tests for new edge cases.