Originated as a forum post

Let's add a tutorial demonstrating how to serve trained models in C++ using torch.jit.save() and torch::jit::load(). I believe there is some subtlety requiring a torch.jit.trace() to wrap poutine.trace and poutine.replay logic for SVI models and guides.

Note that in some prediction tasks, only the guide is needed and no Pyro logic is required. In this case, if the guide is an nn.Module, it can be easily torch.jit.save()ed and torch::jit::load()ed without any need for Pyro.