Per parallel branch agent configuration

Given a large use of parallel is to run chunks of steps on different agents, it may make sense to allow agent to be configured in a parallel branch vs only at the parent stage level.

Currently parallel lives at the steps level, so this likely could be a disruptive change (maybe not).

Currently to use distributed builds in parallel vs sequential means dropping in to using the node syntax (and excludes usage of things like docker inside), which is a bit of a disjointed experience.