Bug fix

• Fix CUDA is reinitialized in subprocess when multiprocess is used for distributed training.
• Add argument device with default value torch.device('cuda:0') to core.get_inception_feature which can be assigned by **kwargs for all utility functions utils.*.