hibernate on critical-load

[shlomi-noach]

Fixes #418

This PR introduces --critical-load-hibernate-seconds. This flag defaults 0 (in which case the new feature introduced in this PR is disabled)

when critical-load-hibernate-seconds > 0, then the behavior of critical-load changes. Instead of panicking and bailing out with error, gh-ost enters hibernation.

As opposed to throttling, hibernation implies absolutely no queries on neither master not migrated server. No heartbeats; no throttle checks; no replication lag checks; nothing.

When gh-ost awakes from hibernation, it's all back to normal, and throttling checks take place as usual, and of course it may potentially re-enter hibernation if critical-load is still met.

cc @ggunson @github/database-infrastructure

[shlomi-noach]

TODO: on leaving hibernation, gh-ost has no up-to-date metrics. It shouldn't just resume work; it must first run a collection of throttle-metrics, inject heartbeat etc. Only then should it resume work.

This is somewhat complicated to achieve and supervise, since most of these processes are independent.

[shlomi-noach]

I believe https://github.com/github/gh-ost/pull/424/files/3955a6d67f80404d1f2611b82023ba4daeda423c..ad47f7c1477780020631627514d3ef168e8f71ff#diff-6298c3b05013fa1f1cc736d61051b34eR292 should do well.