fix bug in optimizer sequence that make the op CrossentropySoftmaxArgmax1HotWithBias and GpuCrossentropySoftmaxArgmax1HotWithBias not being inserted used on the gpu. This caused stability and slowness on the gpu.

CudaNdarray.__setitem now accept an ndarray as a value. It copy the data direction. Restriction, no broadcasting, the CudaNdarray slice must be c contiguous.

optimized memory usage on the gpu: when used gpu_shared_var.value = ndarray, we keep only one memory region on the gpu.

attrs is unhashable since its a list. Fix is to first convert attrs to string
than hash the string.

make the GpuElemwise return the same error as Elemwise when the shape mismatch. Modif test for this.

GpuElemwise now raise an error when we could broadcast input to make the code run but did not set the broadcast flags to True. Added test that we raise this error.

This allow to fuse more op in GpuElemwise(gpu have a limit in the number of fct inputs) and make the graph a little simpler.

Make easier the transition to MRG_RandomStreams from code using RandomStreams
or SharedRandomStreams.