Port Softmax kernel to OpenCL

All tests in test_nnet.py pass with CUDA.
Only fp32 tests in test_nnet.py pass with OpenCL. GpuFromHost doesn't work with fp16 or fp64.
Larger work item size doesn't improve performance.
Add 2 local_barrier(), it's strange that AMD card doesn't need these local_barrier(), but they are necessary for NVIDIA cards.

Fix link to 2014 presentation

(small fix) Re-allow runtime algorithms for conv3d Ops.

Function that check the axes

Jenkins python3 add fastrun and gpuarray

Fix doc typos and spelling errors. Make spelling of "optimize" consistent.

(small fix) update usage of `dnn.enabled`

max work item size on my ocl device(not too old) is 256

Wrap Op params for theano.sandbox.rng_mrg.mrg_uniform

Make dnn_present() always return a Boolean.

Add optimizers o1, o2, o3 and o4, where o1 is equivalent to FAST_COMPILE and o4 is equivalent to FAST_RUN. See #5762

Make version as dev version and better error.

[doc/requirements.inc] Specify which 'lib' or 'lib64' should be added…

When reading this documentation, I first added '/lib' and '/lib64' in LD_LIBRARY_PATH.
In order to avoid other users to do this mistake, I propose to update the
documentation like this.

Make the current version as the dev version. This will make it clear from the version if this is the last release of the dev version.

Convert Github readme.txt to .rst

Faster topo

Fix memory leak in GpuCAReduceCPY

don't move op that don't support float16 on the GPU

(crash fix) Fix availability of cuDNN enumerations in v5 vs v6.