expose retain_graph to user within TorchAgent

[prajjwal1]

Patch description
Although loss.backward() is the norm in usual workflow, there are many instances in which a user might want to pass in retain_graph=True within loss. Considering that parlai doesn't expose the loss api to user, since TorchAgent expects user to call self.backward(), it would be a good addition to let the user pass in retain_graph=True if they want to. User can also pass in create_graph=False, inputs=None params if they want to, to backward().

Testing steps

Other information

[klshuster]

lgtm