Improvements to DID

[drbenvincent]

• Partial progress on DiD: allow for multiple pre and post intervention observations #76. Makes DID analysis more robust. The initial implementation made too many assumptions making the code brittle and not generalisable to other datasets. It also assumed that there were multiple units per condition (ie multiple control and treatment units). This code is much more robust and now works when there is a single treatment group. That said, more work needs to be done in this case.
• Adds a 'classic' example, the bank failure dataset, so contributes to Add examples for 'classic' causal inference datasets #44. So far this is only implemented with the Bayesian model, not the scikit-learn model.