The old version took the first job in the scheduled (or retry) set,
tried to schedule it, then looped.

The new version takes the first 100 jobs from the set, tries to schedule
each in turns, then loops. This is designed to reduce scheduling
latency, but on GitLab.com we're seeing increased Redis CPU usage and a
high number of ZREM commands (most of which will be redundant) due to
multiple poller threads from different processes operating at the same
time.

This change reverts to using the old behaviour, while remaining on
Sidekiq 6.