Scaled Normal Noise Model

[SamWitty]

This PR addresses an error when conditioning on data using the ScaledBeta noise model.

[anirban-chaudhuri]

The scaled Normal implementation now looks good to me.
However, I had a couple of other comments:

• When the observation is 0, should the observation model use the scaled normal by default? If so, it could be added in the if-statement that currently exists.
• Otherwise, an observation of 0 will lead the scaled beta distribution to have alpha=0, which will not work. Am I understanding it correctly that you are setting alpha=scaled_mean * pseudocount and beta=(1 - scaled_mean) * pseudocount for the beta distribution (or, I might be misunderstanding the pytorch description for the inputs)?

[anirban-chaudhuri]

Looks good to me.
Issues we might need to revisit:

• Scaled beta distribution for very high and very low values of alpha and beta parameters for Beta distribution
• Truncated distribution instead of Scaled Normal to avoid negative values and values greater than total population for states in covid models