Checkpoint/migration handling

[MingweiSamuel]

Split out of #929

Let's get all the state of a transducer exposed for management including checkpoint/restart.

• What about state that persists across executions (e.g. in storage or a replication library)?
• Relates to the "tables" in Bloom being absent from Hydroflow: is state first-class in the IR and separate from ops?
  • Might be good to have an explicit data model spec

from @jhellerstein

[jhellerstein]

Arguably our philosophy could be that the Hydroflow runtime itself does not provide intrinsic mechanisms for efficient checkpoint. Instead we defer checkpoint to a system either beneath Hydroflow, or built out of Hydroflow.