Reorganize Edward tutorials

[dustinvtran]

@dblei argues that we should reorganize the tutorials. Namely, a tutorial should be about a single topic, such as analyzing a specific data set, model, inference, whatever. A reader with basic background in probability modeling should be able to read the tutorial from beginning to end—without resorting to other pages—and understand how to address that topic with Edward.

Old design

New design (with fake tutorial headlines)

Remarks

• From the actual list of tutorials available, it's clear that Edward needs more tutorials teaching inference and criticism. We also need more real data set experiments.
• The Bayesian neural network and probabilistic decoder tutorials are incomplete, in that they do not apply the model to a data set and perform inference/criticism to learn hidden structure. These will be opened as Github issues after this PR is merged.
• Inference and criticism tutorials should explain how they're implemented. This mimicks how model tutorials are done, when we build them up with random variables. Most importantly, this gives a more realistic depiction of Edward, which is that it's not about applying built-in tools. Rather, it's about experimentation and possibly building your own tools.