Add cuDNN v5 support, drop cuDNN v3 support

[flx42]

cuDNN v5 is now out, since PR #3919 is still being discussed, we need a solution to allow users to compile caffe with the new cuDNN version.
This PR is the bare minimum to enable cuDNN v5 with caffe while still being compatible with cuDNN v4 to give some time for users to transition to the newer version.
cuDNN v3 support is dropped to avoid further #ifdef hell.

Ping @longjon @shelhamer @ajtulloch

[thatguymike]

Note that without the memory management changes and some of the commits to use it to enable more algorithms in cuDNN, performance is still far from Nvidia's caffe branch for most popular networks.

[shelhamer]

Thanks for the minimal patch for compatibility @flx42! Thanks for noting the difference with further integration of the new algorithms in v5 @thatguymike.

With merge I will post the version change to the mailing list and publish the latest version of the Caffe site to incorporate the doc change.

We will turn to the full integration of cuDNN v5 after the NIPS deadline 05/20.

[Coderx7]

what is the status of cudnn v5 now? Is it still WIP or is it finally dealt with like nvidia branch?

[shelhamer]

Caffe has been cuDNN v5 compatible since this PR was merged. However, certain memory hungry convolution algorithms are more practical with CUB memory pool integration, which is still in progress.

[lld533]

In Line ~40 of cudnn_{tanh, sigmoid, relu}layer.cpp, to prevent possible memory leak, it's better to destroy activation descriptor by calling "cudnnDestroyActivationDescriptor(this->activ_desc)"; before destroy the handle_. Much thanks.