Model checks

[conorheins]

This PR:

• adds dimensionality and normalization checks in the constructor of the Agent() class and gives informative error messages when things don't line up (either certain distributions are not normalized, or dimensions don't line up)
• simplifies input structure to Agent()
• certain variables like num_controls are automatically inferred from structure of user-given B array, if not provided.
• control_fac_idx is tested more rigorously against other model dimensions (like the B array and num_factors)
• extra unit test for verifying that observations with arbitrary structure (tested: one-hot, tuple, list, int, and object array of one-hots) can be passed into learning.update_likelihood_dirichlet(...), based on an issue raised by @daphnedemekas

[conorheins]

@daphnedemekas : can you verify that the update_A works in your setup with arbitrary observation structure? Let me know if you still were having that issue where you had to reformat your observations, and I'll fix it. You shouldn't have to hack around it manually - the utility function process_observation is supposed to be able to deal with arbitrary observation structure.