Adapting discovery demo for solubility task

[jannisborn]

• Solving Adding more components to the target scoring function #181: Improving PaccMannGP to support arbitrary callables as suggested in the discussion
• Adding a new model PaccMannVAE that produces unconditional molecules. It's the same decoder as in PaccMannRL but just used by sampling randomly. We basically just wrap the PaccMannRL model.
• Adapting the case study on GENTRL's DDR1-inhibitor to optimize solubility instead of QED. Reran all the models, including the new one (PaccMannVAE). Findings are the same in that MoLeR and RT are performing the best

[drugilsberg]

The logic looks good, I did not have a chance to run the full notebook, but the logic and the updates look good.

[drugilsberg]

@jannisborn small build due to a PaccMannGP test:

if self.mode == "max":
[5965](https://github.com/GT4SD/gt4sd-core/actions/runs/3853838762/jobs/6567211532#step:9:5966)
            return 1 - (sum(values) / len(values))
[5966](https://github.com/GT4SD/gt4sd-core/actions/runs/3853838762/jobs/6567211532#step:9:5967)
        elif self.mode == "min":
[5967](https://github.com/GT4SD/gt4sd-core/actions/runs/3853838762/jobs/6567211532#step:9:5968)
>           return sum(values) / len(values)
[5968](https://github.com/GT4SD/gt4sd-core/actions/runs/3853838762/jobs/6567211532#step:9:5969)
E           ZeroDivisionError: division by zero

Looks like it is something dependent on the error handling in the scoring function in the package.