Improve MC-AIXI Performance & Ensemble capability

• Improve performance for MC-AIXI, by reducing complexity of the RateBackend AIXI adapter
• Fix an issue with tie-breaking -- this enforces MC-AIXI alignment, previously we used the fastest first, now we (pseudo)randomize it. No more MC-AIXI bugs.
• Add Convex Mixture from "On Ensemble Techniques for AIXI Approximation" to complete implementation of the paper. Also updated the Switching mixture logic to match strictly with that paper as well.
• Planner backends were generalized so both MC-AIXI and AIQI can use the shared RateBackend abstraction directly
• Improve tests