Backlog

There are a few things that we went with because they made sense at the time that I want to double check again. This issue is a place to collect them.

normalize_axis in index functions. Is it really necessary to accommodate negative axis arguments and is this the way to do it?
Do we need this strong separation of dask and numpy implementations? After some changes (perhaps both in numpy and dask) it seems that now even numpy functions will generally return dask arrays if presented with such.