Greentea LibDNN benchmarks

[naibaf7]

This PR enables LibDNN as the default convolution engine in Greentea benchmarks.

Note that LibDNN is available for both CUDA and OpenCL, and I also enabled this as default in the Makefile.
It is important to select the correct GPU: I have set GPU 1 as OpenCL and GPU 0 as CUDA in the benchmark scripts, but if the system has more than one CUDA GPU, then the GPU ID for OpenCL has to be changed to #CUDA GPUs + 1 (i.e. 2 if 2 CUDA GPUs are present).

@soumith
Could you do me a favor and benchmark this in CUDA and OpenCL? Would help a lot in the further development of the convolution engine.

It is not yet autotuned for the Titan X but should give a hint on performance anyways. If the OpenCL score greatly differs from the CUDA score, this is most likely due to the slow ViennaCL operations used for weight updates and auxiliary math operations. CLBlast and clBLAS would do better, but compiling them takes more time.

Thanks :)