ENH: add 'pretrained' and 'weigths_path', fix #87

[NickleDave]

adds ability to specify whether model should be pretrained, and to specify a path to pre-trained weights that should be loaded

• add 'pretrained' parameter to TransferTrainer.from_config and Trainer.from_config classmethods
• add 'pretrained' attribute to config.TrainConfig
• add 'pretrained' parameter to searchnets.train.train
• have main call train with parameter 'pretrained'
• fix parse so it passes 'pretrained' to TrainConfig
• add 'weights_path' parameter to all model build functions
• add 'weights_path' parameter to 'from_config' methods for Trainer and TransferTrainer
• add 'weights_path' attribute to config.TrainConfig'
• fix 'config.parse' to unpack WEIGHTS_PATH option and pass to TrainConfig
• add 'weights_path' parameter to searchnets.train
• have main pass 'config.train.weights_path' to train
• fix how model build functions unpack state dict from checkpoint
  • check whether checkpoint contains a state_dict key, if not, assume checkpoint itself is just state dict
  • remove 'module.' from state_dict keys to avoid crash -- occurs when checkpoint was saved during training with nn.DataParallel