Add bench tests to track performance improvements

[ehiggs]

Add bench tests to track performance improvements.

[MichaelHirn]

The beneches for all native available methods got introduced with #7. Benches for CUDA (cuDNN) are still missing though.

[MichaelHirn]

Update, since the introduction of Leaf 0.2.

Leaf has a debug functionality, that can be activated when executing the examples, which prints the execution time for each layer. The Leaf layer has some overhead, but it can be taken as a proxy for the "raw" operation e.g. convolutions via cuDNN.

That being said, this issue is not closed yet, as we would like to have it more isolated here in the library.