CI: Simplify and unify environment creation

[datapythonista]

When creating a pandas environment for the CI, we currently have two separate systems, one for Travis and one for Azure. Summarizing, those are the involved steps.

travis

ci/prep_cython_cache.sh
ci/install_travis.sh         -> locale, download/install conda, ccache, create env, setup.py develop
ci/submit_cython_cache.sh
ci/install_db_travis.sh      -> create dbs
ci/before_script_travis.sh   -> xvfb (clipboard)

azure

ci/incremental/install_miniconda.sh          -> download/install conda
ci/incremental/setup_conda_environment.sh    -> create env, locale
ci/incremental/build.sh                      -> setup.py build_ext

Except for the cython cache, I think all the rest could be unified in a single file ci/setup_env.sh that performs all the steps.

This would simplify and avoid duplicate code and inconsistencies. As well as making future changes simpler.

[saurav2608]

@datapythonista - Can I take a shot at this?

[datapythonista]

sure, go ahead

[saurav2608]

@datapythonista - I see two different build commands. In ci/install_travis.sh the command is python setup.py develop || exit 1. And in ci/incremental/build.sh the command is python setup.py build_ext -q --inplace. What is the difference between these two?

[datapythonista]

Afaik, setup.py develop installs the package locally (compiling and with dependencies), setup.py build_ext is just the compiling part.

Not sure about the reasoning here, may be things changed in the setuptools/pip since this was first set up. I think setup.py develop is not used much nowadays. When not using conda, pip install -e . is usually the preferred way, as pip is smarter than setuptools with dependencies. If we're using conda, I think setup.py develop should be unnecessary, but may be there is a reason I don't know in this case.