don't drop by half periodically.

[abersnaze]

after talking with @jolynch this section doesn't make sense if the next line bounds the limit between the min/max. #155

mentioned in #72

With AIMD, this works moderately well - eventually the client goes above the rate limit and a request is dropped. The next window, the concurrency is dropped by half, which generally will correspondingly drop the concurrency too, and cause the rate limit to be missed.

tangentially related to issue #147.