Group updater: add support for parallelism

[wfarner]

Addresses this TODO in rollingupdate.go:

// TODO(wfarner): Make the 'batch size' configurable.

We should accept a parallelism flag for updates to allow multiple instances to be drained simultaneously.

[chungers]

I wonder where this will go. Is this a responsibility of Flavor or the group? I would think the Flavor worries about how to do a drain, and Group worries about how many at a time -- assuming different instances can be drained in parallel.

It's the inverse of the instance's Provision. So in the sense that a Flavor's Prepare is chained to the Instance's Provision, a Flavor's Drain should be chained to the Instance's Destroy.

Related #235

[wfarner]

Flavor's Drain should be chained to the Instance's Destroy

That's exactly what i was thinking.

There are some use cases that we should anticipate (but probably not solve yet) - where parallelism should be selective. For example, in a Swarm we would ideally not pick all N instances hosting a particular service.