Various small changes due to the chamnges in the semantics and syntax of the
PyTorch 0.4 API.

Note that currently distiller.model_performance_summary() returns wrong results
on graphs containing torch.nn.DataParallel layers.