Zero-rate edges confuse distributed mode

[jbosboom]

If we split the graph across a zero-rate edge, the input TokenActor in the downstream blob will report rate 1. Because the downstream actor has input rate 0, the only valid schedule is to not execute anything at all.

Ideally we would determine the TokenActor's output rate by the upstream actor, but that appears difficult after the graph has been partitioned (because the upstream actor is not in the partition). Instead, we can use the simple heuristic that the TokenActor's output rate is 0 when the downstream actor's input rate is 0, and 1 otherwise.