Support models which produce [co]-variance estimates

[WardLT]

We should add supports for algorithms which generate both prediction means and uncertainties (e.g., GPR)

Our selection algorithms operate on samples from the prediction distribution. The easiest way to integrate ML models which produce their own prediction distribution is to draw samples from them. To do so, we must:

• Provide an option to generate samples from the means in the Scorer
• Allow the Thinker to recognize when the Scorer has generated multiple samples for a single model
• Support the generation of multiple samples in the RDKit scorer (at least)
• Implement an example