Before this change we read sleepTime before acquiring the pacer token
and uses that possibly stale value to schedule the token return. When
many goroutines enter while sleepTime is high (e.g., 10s), each
goroutine caches this 10s value. Even if successful calls rapidly
decay the pacer state to 0, the queued goroutines still schedule 10s
token returns, so the queue drains at 1 req/10s for the entire herd.
This can create multi‑minute delays even after the pacer has dropped
to 0.
After this change we refresh the sleep time after getting the token.
This problem was introduced by the desire to skip reading the pacer
token entirely when sleepTime is 0 in high performance backends (eg
s3, azure blob).
83d0c186a7