Trim dependencies (and make consistent)

[mortonjt]

We definitely don't need

• sklearn
• seaborn
• ipython

Need to trim before version 1.0

Also python3.7 is breaking for some reason ...

[fedarko]

Looks like seaborn and ipython aren't used anywhere in the current codebase, although sklearn is (we're importing check_random_state from it). Weirdly enough the setup.py and conda_requirements files don't mention needing sklearn... I guess everyone has just been installing into environments where sklearn is already present :)

This relates to something I've been thinking of for a bit -- we should ideally try to make all of the dependency info consistent. There are (at least) three locations where this info is specified for Songbird:

1. setup.py
2. ci/conda_requirements.txt
3. conda-forge meta.yaml recipe

And all of these are slightly different in some ways -- e.g. setup.py has a different min numpy version than conda_requirements, and the conda-forge recipe explicitly mentions needing sklearn while the other two don't.

Ideally we should be able to make this info consistent somehow (would it be possible to just delegate everything to the setup.py file?), or at least periodically check in to make sure stuff's in sync.