Add SWA support to Pytorch trainer

[andrewyates]

Pytorch 1.6 supports stochastic weight averaging (SWA), which could improve model effectiveness. The Pytorch trainer should support a boolean swa config option that controls whether it is used.

As described in the docs, using SWA with batchnorm requires additional steps. Can we automatically detect whether a model contains any batchnorm layers and, if so, either raise an exception (easier) or perform the necessary additional steps?