Smarter reference management for the scheduler

[vsreekanti]

Right now, the scheduler deserializes all arguments and looks through them in order to determine which ones are CloudburstReferences and which ones are not. This can be really expensive if deserialization is slow and/or the data is large. Instead, we should elevate these references into the protobufs, so the scheduler can look at them without doing any deserialization. Note that we will still want to keep the references in the arguments themselves, too, because we need to know where to re-insert the references once we resolve them.