Some http improvements.

- Attempt to re-use connections when we've already received enough
   data to do so immediately.
- Make sure when seeking near the end, if the current chunk size is
   such that the _next_ chunk will be half the normal size or less,
   we just ask for the rest of the resource.

With these two changes, a normal open of a single-chain Opus-only
 file requires exactly two HTTP requests.

- Also use the response buffer as a dummy buffer when skipping
   data.
  This will avoid helgrind errors for multiple writes from
   different threads without locking (should someone be reading
   multiple streams from different threads).
  It's also better for SMP cache contention.