Add support for no indices in advincsub1 for gpuarrray.

Crash Fix

pool.py optimization with openmp

[FIX TRAVIS] Make the new node on a graph not depend on the old node.

[WIP] Use check_stack_trace helper function in tensor/nnet/tests/

Make get_scalar_constant_value investigate Second as Alloc (so elemwise=False still check it). This fix tests in travis in FAST_COMPILE.

Removed repetition

Update Docker instructions

NaN Introduced by AllocEmpty

Split test to help work around travis timeout. They are super fast here.

Don't rebuild inplace add kernels all the time for GpuIncSubtensor.

[ENH] faster opt by changing call to extract_constant and get_scalar_constant_value

A few other fix to make the SBRNN code compile in the new back-end in fast_compile

This reverts commit d4965495.

This reverts commit ea3e6137.