Re-add connection retries, but less aggressively.

In the course of #863 connection retries have been removed, because the current logic seems a bit dangerous. What we want is roughly something like:
- perform at most three retries in a single transaction, then fail. That prevents hundreds of retries when e.g. COPY is used, making the problem worse
- add some delay in retrying e.g. connection_timeout / 10 * num_retry or such
- don't retry if the error is due to max connections