PartitionCache for DynamicSchedule

[xcharleslin]

Currently, the partition dependencies of a DynamicSchedule reside within the DynamicSchedule itself. This makes sense for ownership, but is a bit smelly of an interface as the partition constructor must reach into the DynamicSchedule to populate its dependencies.

A better approach would be to extend the existing PartitionSetCache concept to handle individual partitions and use this for DynamicSchedule. Then, partitions can be constructed agnostic of their eventual owner(s).

[xcharleslin]

Resolved the ownership issue a different way in #434 (ExecutionSteps receive a result accessor; runners can implement their own result accessor) and avoided having to deal with any explicit partition management for now.

Hopefully this will suffice. I think it would be ideal to rely on reference counting and avoid explicit global partition management. In the future, individual runners might have to implement some partition management; ideally this would remain properly abstracted away from the rest of Daft.