Tasks where placement is SPREAD_ALL_SLAVES should not be shuffled

[relistan]

If you currently have both shuffleTasksForOverloadedSlaves: true and spreadAllSlavesEnabled: trueenabled, and have requests running with SPREAD_ALL_SLAVES placement, then having any busy slaves currently results in a ton of task churn for no reason and ends up causing lots of poller delays and task instability.

I believe a check needs to be added here to prevent Singularity from trying to move tasks that have been added to the slave by the request placement SPREAD_ALL_SLAVES.

It looks like the current workaround is to specify those tasks as non-eligible manually. But the system should figure this out on its own.