zephyr: Use exponential backoffs in retry loops.

[alexmv]

This reduces the number of retries that might spam APIs.

There is some complexity here which is left un-managed -- for instance, maybe_restart_mirroring_script does a number of restart attempts, and then fails, but will be retried every 15s by the surrounding process_loop. Previously, it would merely have looped forever inside maybe_restart_mirroring_script.

Three loops are intentionally left as infinite while True loops, that merely cap their backoff at the default 90s. Their callers do not expect, or have any way to handle more gracefully, a failure of the expected-infinite-loop in process_loop or zulip_to_zephyr. They maintain their previous behavior of retrying forever, albeit more slowly.

[alexmv]

See also #586.

[timabbott]

Posted some comments; otherwise lgtm.