Fix model_optimizer_options argument

[dmitriy-serdyuk]

Either I do not understand how to use option model_optimizer_options or it will not work with current implementation.

This PR allows to write --o.model_optimizer_options "encoder={'lr': 0.01}".

[dmitriy-serdyuk]

Actually, why do you need this argument? I can see now that that --o.optimizer_options supports passing a dictionary.

[rdevon]

@dmitriy-serdyuk Can you rebase this to dev? I think there might be some overlapping in the branches

[dmitriy-serdyuk]

Done

[rdevon]

@dmitriy-serdyuk OK good. I'll fix the flake errors with dev and start merging.