Implementing `groupby`

Incomplete notes (most is still on paper or in headspace):

• Need to support mapped nodes in __setitem__ first, otherwise groupby is relatively useless.
• Figure out what it means and how to handle the node mapped for grouping: When grouping a DataArray (or DataFrame) we have a coord param before grouping and another coord param (with unique param values) after. How can the be reconciled with the graph, where we can't have a node twice? Or does it work already, because MappedNode includes the index name, so there is no conflict, except that convenience features like node-name based lookup do not work anymore?
• There are cases than can be drawn easily, but are hard to setup using the syntax similar to map/reduce, combined with __setitem__ to continue the graph after reduce. Not clear if those are needed/useful, but maybe the current syntax is not sufficient.

Start by reading the existing map/reduce syntax, setitem, and MappedNode handling. Clarify how mapped grouping nodes and coordinates should behave, including whether graph lookup remains usable and whether post-reduce continuation is required; done means these semantics are agreed and groupby is implemented accordingly.