Add 'grace time' behavior to Sidekiq MemoryKiller

The previous iteration of the memory killer did not give 'large'
jobs much of a chance to finish. This caused them to come back after
each restart, leading to another memory-induced shutdown. By adding
some 'grace time' between the moment we detect a memory peak and
the moment we start shutting down, we hope to give Sidekiq a chance
to get whatever causes its memory to grow 'out of its system'.

See merge request !1313