Add PredictionsOpt, fixed restrict_models bug, code cleanup

[Innixma]

Add PredictionsOpt, fixed restrict_models bug, code cleanup

• TabularPredictionsInMemoryOpt is slightly faster than memmap and operates in-memory.
• Fixed restrict_models bug in memmap that caused the model predictions to be returning the wrong model results after the restrict call.
• Various code cleanup, added logic to specify the predictions format to load with in Context via prediction_format.
• Added extensive unit testing to all TabularPredictions classes, including equivalence testing.
• Added logic to have all TabularPredictions classes be able to load from data or dict, and return dict, and save to data.

[Innixma]

Added benchmark_repo.py to show a comparison of runtime. Note that for longer zeroshot simulations, the gap will increase and eventually be larger than the gap of load time. Regardless, both memmap and memopt are close in speed while being much faster than mem. I think it is good to support both, with memmap being the default in most cases.