Allow pytorch model to customize collate_fn to be used in torch DataLoader

[eddyxu]

All the official pytorch object detections model expect the input as a list of tensors List[torch.Tensor[3, H, W]], while the classification models expect a mini-batch of stacked tensors torch.Tensor[N, 3, H, W].

Object detections model then can do resize in GPU and tolerate different size / channels of images.

We could probably expect that the none-official models will have a wide range of expectations from collate_fn. Rikai should be flexible about that.

[eddyxu]

By default, Torch DataLoader will throw error with SSD + the full coco dataset even we apply Resize(300, 300) to the images.

[eddyxu]

Closed via #643