Allow aggregate window transformation to operate with multiple predecessors

[jsternberg]

The aggregate window transformation currently only allows one predecessor. It will take an input table chunk, compute an internal state for that table chunk, merge it with an existing state (if one exists), and then it will compute a final output when all of the inputs have been consumed.

We want to modify the transformation so it is able to do this with multiple predecessors and to then allow itself to be used with a parallelization factor so it can aggregate chunks from its parent concurrently.

DOD:

• Aggregate window can take inputs from multiple predecessor datasets and then merge them.
• It exposes planner attributes that identify it as a merge node.

[jsternberg]

Fixed by #4851.