General groupby optimizations

[tedmiddleton]

• Would sorting the row indices make grouped_frame faster or slower?

• strings are still quite slow in mainframe. I have to investigate whether there are more opportunities to optimize copies in series_vector as well as perhaps implement string hashing in series or frame.

• just thinking about this, mainframe could feature a transformer a bit like frame::allow_missing() but would instead convert a std::string column into some other sort of thing that uses hashes. This would sort of mean making a hashed string class. Maybe even ref-counted? Something that will coerce easily to and from std::string?

• Could series_vector be even more optimal in its copies/moves when dealing with primitives and pods?

• even unordered_map seems to be a bit slow for what I'm using it for. I wonder if that could be improved?