Refactoring Tabular MDP code

[markkho]

Organizing some thoughts on how to clean up tabular MDP code for a v1.0.0 release (@cgc)

• Streamline testing with a standard set of benchmark tasks and comparison of outputs for exact algorithms
• Create a modified policy iteration implementation that has value iteration and policy iteration as special cases
• Include a non-vectorized version of MPI/VI/PI for easier debugging
• Reorganize msdm/core directory structure to be flatter
• Explicitly define MDPPolicy since these will be different from POMDPPolicies, etc.
• Streamline translation between different policy and state/action mapping representations (e.g., matrices vs. dictionarys), and html-table-like output for jupyter notebooks
• change mdp run_on to return a trajectory of Step objects defined by namedtuple (similar to what pomdp does now)

[markkho]

starting to address this in #65