Discussed code re-organization

[birdsarah]

Fixes #38, #37

• clean-up requirements files
• make python setup.py test actually run the tests
• roll flake8 testing into every test run
• move flake8 config into one place
• I did not port the two tests that were added here: https://github.com/mozilla/taar-lite/pull/32/files#diff-c791252c63cf945e8e08b32e0689432fR199 - that were part of a fix for #31. This issue can't happen any more because things aren't computed at different times. Each entire recommender is re-setup when the files are re-loaded.
• tweaked travis to run python setup.py install so that path gets tested a bit
• added a MANIFEST.in for packaging
• renames production.py.GuidBasedRecommender to TaarLiteAppResource
• removes guidception from production
• moves guidception into experimental treatments
• ports over all tests (except two mentioned above), simplify pytest fixtures and test code where we can now just test standalone treatments and recommender
• tests that production recommenders are configured correctly (rather than testing their output)
• move _strip_low_rankings out of recommender and into treatments and add concept of multiple treatments to recommender

Marking as fixes #37, because guidception has been moved to experimental.

[birdsarah]

Tests are failing because dockerflow is not a requriement of srgutil, which is where the logging comes from. I believe that's where the dependency should live. Thoughts @crankycoder?

This has been addressed and updates pushed to srgutil.

[birdsarah]

@crankycoder, pending the test failure, which I will fix, this is ready I think.

There's a couple of API niggles (e.g. I'm thinking about tie breaking as a step like treatments). @dzeber is working on a follow-on PR to:

1. clean-up treatment code
2. add metrics to recommender class

@dzeber, you'll see i've added an analysis directory and some instructions on README. Contrary to what I said earlier you can start Jupyter from wherever.

cc @mlopatka

[birdsarah]

@crankycoder I would probably do a squash merge due to so many incremental (and failing) commits along the way here. (But I generally prefer squash merge, so up to you)

[crankycoder]

Sorry for the long delay in completing the review. It wasn't clear to me that adding the treatment graph code modified the range of addon-ids that taar-lite would respond to.

For other people's sake - I used the following URL when running taar-api-lite to test end to end functionality.

http://localhost:8000/taarlite/api/v1/addon_recommendations/html5-video-everywhere@lejenome.me/

The previous GUID I happened to be using "{a0f88751-df67-4902-8da4-543cc4ce3d48}" wasn't in the treatment graph so I was surprised to see an empty list of results come back.

[crankycoder]

It's friday. I'm just going to merge this in.