Estimate size with arbitrary functions in `.forward()` using graph history

[jacobkimmel]

The original implementation of this tool relies on accessing tensor operations through model.modules(). This is simple, but cannot account for arbitrary dimensionality changes in the .forward() method. For instance, it's common to perform a tensor.view(...) or torch.nn.functional.max_pool() in .forward().

We could alternatively extract the operation history from the graph of functions recorded by autograd. See https://github.com/szagoruyko/pytorchviz/blob/master/torchviz/dot.py for an example where the graph is extracted.