Add a method to initialize dynamic modules from state_dict

Hi,

**Training**
I initialize `net=MTSimpleCNN()` and train this multi-task model with multi-head classifier on two tasks (one task per experience). After training the model has two heads.
I save the model this model's state dict using standard pytorch `torch.save({'state_dict': net.state_dict()}, filename).` and I want to evaluate it later using a different script.

**Evaluation**
I newly initialize the model `net=MTSimpleCNN()` and load the weights using standard pytorch `net.load_state_dict(torch.load(filename,map_location=device)['state_dict'])`. However the saved model state dict has two heads and the newly initialized one only one head. Keys don't match so I get this error:

`Error(s) in loading state_dict for MTSimpleCNN: Unexpected key(s) in state_dict: "classifier.active_units_T0", "classifier.active_units_T1", "classifier.classifiers.1.active_units", "classifier.classifiers.1.classifier.weight", "classifier.classifiers.1.classifier.bias", "classifier.classifiers.0.active_units"`

is there any easy way to initialize this multi-head network with arbitrary number of heads?

Cheers,
Woj