add ability to specify whether model should be `pretrained`, and to specify a `weights_path`

[NickleDave]

to be able to load models trained e.g. on different datasets besides ImageNet, or to perform transfer learning on a randomly initialized model as a control