Support ResNet

[yiheng]

We need to support resnet in spark-dl, including single node performance test and tunning. Test convergence on single node and multi-node.

We can start with ResNet50. The model topology should reference facebook torch model and caffe torch model. The only difference between these two models are some convolution stride parameter.

The single node performance goal is to achieve 80% of intel-caffe.

The convergency goal is to achieve the same error rate(top-1 and top-5) on the refrence model.