Created by: qingqing01

This code are splited into pull https://github.com/baidu/Paddle/pull/218 and pull https://github.com/baidu/Paddle/pull/219 . SO, please to review them.

1. add benchmark
2. add ConvProjection to reduce GPU memory for googlenet
3. increase workspace limit size in cudnn_conv to speed training.
4. use TmpMatrix in cudnn_conv to reduce GPU memory.