ADR: architecture for fast groupby behavior

[machow]

1. When should it fail, vs fall back to apply? Should it emit a warning? How could this be customized?

2. How could singledispatch be used with CallTreeVisitor? Can we demo using a vector function (e.g. row_number, n, cumsum)

3. Could type annotations be leveraged here? E.g. automatically wrap a vector function based on it's return type, or some other function annotation.

4. How can a user audit the behaviors siuba will take before executing e.g. a grouped mutate?

5. Premortem: pretend we're looking back 3 months from now, and the process is not working--what are the main reasons that is (e.g. user perspective, dev, perspective, person who has been programming for 1 day perspective)?