Rewrite calls to assert_allclose() as suggested by @nouiz .

(to help specify expected value vs given value)

Update documentation and config flags about supported cuDNN algorithms.

Don't print float16 warning for ops that don't have c code

jenkins buildbot with docker

Add lifter for CrossentropyCategorical1Hot and grad

Added mode 'half' to Images2Neibs. Tests pass.

[ENH] Speed up nanguardmode by not checking *int* dtype

(small fix) Add doc for `dnn.include_path` and `dnn.library_path`.

Split long-running test so that it helps travis not to give up

Fix offset problems in the new backend.

Fix issue in optimizations with bitwise operations

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.

Added checks in the GPU code for invalid inputs.

There is no corresponding GPU Ops, but we can express them
as combinations of GPU Ops.

0 or 1 are not converted to boolean types if that is what
the user would want.

This also reflects the difference between bool and int8

Since tensor.or_ and tensor.and_ are bitwise operations
(for non-booleans at least), some "logical" simplifications are wrong.

Fix link to 2014 presentation