Allow custom distributions in policy gradients

[befelix]

Not sure if this is consistent with other parts of the codebase, but this is how I currently provide custom action distributions for policy gradients:

actions = dict(continuous=True, distribution=Gaussian())

where Gaussian() can be replaced by any instance of Distribution.

[AlexKuhnle]

Thanks @befelix for the PR! I added it and will make it consistent to the codebase.