[metalearn] support arbitrary expansion of action space in metalearn_controller

[cosmicBboy]

It would be helpful to arbitrarily expand the number of actions that are available to the metalearn_controller while being able to load weights from a pre-trained controller.

When constructing a new MetalearnController, I want to be able to specify a potentially different algorithm_space such that it contains new AlgorithmComponents and hyperparameters for those components. I then want to be able to load_pretrained(model) so that overlapping nn.Modules are loaded onto the new controller, while freshly initialized weights for new components and hyperparameters are preserved.

[cosmicBboy]

fixed by fb48dc65f210bf275f41650e087ff2801c0de4cc