DecisionPolicy PR

[laerdon]

This PR is meant to introduce decision policies as a flexible layer for turning model scores into final intervention decisions.

Instead of each forecaster handling its own threshold tuning logic, these changes move that work into the shared DecisionPolicy base class.

We also provide an implementation of a deferral-oriented policy that can look at simulated future replies before deciding whether to intervene.

Update:
More recent commits introduce some extra changes:

• Adding in other decision policies, including the RandomDeferral, SimulationAverage, and SimulationMajority decision policies.
• Edited the notebook decisionpolicy_demo.ipynb to further anonymize data, and added section headings.
• Edited docstrings which had not been fully updated.
• Excised unnecessary code in DeferralDecisionPolicy.
• Black reformatting.

Update 6/8:

• Fixed update in init with adding more decision policies
• Added decisionpolicy rst for docs
• Cleaned errant TODO in TransformerDecoderModel.py

Update 6/12:

• Checked package dependencies
• Added link to arXiv in decisionpolicy.rst
• Cleaned TODO from demos (got rid of local loaders)