Remove duplicate tests. No need to retest them all in 2d and 4d. Fusion don't depend too much about ndim

jenkins pr master cache

Compile less C code

Speed up some tests

Try to fix travis. Don't recompile du to new flag when not needed.

Bump min Python version to 3.4 (fix slowness travis in Python 3.3)

Add a DEBUG option for C code

Add deterministic={default,more} flag

Params for other ops

Enable float16 for GpuGemmBatch.

Fix opt crash in float16 and enable C code for MaxAndArgmax and Argmax

Improve usage of op params for neighbours ops (try to fix jenkins error).
Update AdvancedIncSubtensor1.c_code_cache_version (as we change the type of an op param).

Add reduction based on cudnn.

 - i (scalar, int64)