Add support for caffe-style ResNet and more flexible pretrained model loading

[bowenc0221]

• Add a config.MODEL.STYLE field in the config to support caffe-style ResNet (by setting config.MODEL.STYLE='caffe')
• Support more flexible pretrained model loading
  • now support directly loading models trained with PyTorch ImageNet example
  • pretrained model file can be a dictionary with the key: state_dict
  • pretrained model can be directly saved by torch.nn.DataParallel (keys begined with module.)

[msftclas]

All CLA requirements met.

[bowenc0221]

@leoxiaobin

• add caffe-style ResNet results in README
• pretrain models are uploaded in Google Drive, you can download these models and put in your shared folder