PyTorchShim: Add serde callbacks to facilitate lazy loading models

[shadeMe]

When attempting to lazily initialize/load PyTorch models, special care needs to be taken when deserializing trained models from disk. Since the init method of Models are not called during deserialization, shape inference cannot be performed before restoring model parameters. Therefore, the necessary information required to initialize the model must additionally be serialized to disk for later use during deserialization.

[danieldk]

Looks good to me, but since this is a fairly large API extension, someone else should also look at it.

[rmitsch]

Looks reasonable from what I can tell with my limited exposure to thinc. Only minor comments.

[rmitsch]

One minor comment, otherwise LGTM. Approving this to not block things unnecessarily.