package dependency refactor

[bjlittle]

This PR reorganises the conda dependencies for developers.

This is a simplification but also separates concerns between conda and pip.

All pip dependencies are maintained within the setup.cfg. If people want to use pip, then it'll work just fine.

However, as developers I believe that we should be opinionated, and that said, the recommendation is to use conda over pip, or if needed, pip within conda.

The pattern of maintaining the conda dependencies are now all under ci/requirements/py3. These are explicit dependencies per python version. This now allows us the freedom to have varying dependencies per python version, if needed - but in general my expectation is that this typically will be an exception.

The .travis.yml is now pretty simplified, as all the dependency magic is explicitly baked into the pyXX.yml configuration files. I'm keen that if this pattern is accepted that as a rule we can adopt the same approach to other SciTools repos - in particular iris which is a bit of a dogs breakfast when it comes to dependency management, particularly for users and developers alike.

@trexfeathers

[trexfeathers]

I get the principle here @bjlittle, but it's a pretty big violation of D.R.Y.. Is there anything that can be put in place to help developers remember some changes will need to be repeated across files?