Allow to move to the GPU Composite op that have non float32 in them but float32 in input and output.

test that we con't move Composite with float64 in them to the gpu and that we don't move elemwise comparaison to the gpu even if the output is cast to float32.

It removes annoying warning messages from the buildbot's output.

Tests used to fail due to accumulation of negligible errors, so the
differences were big after a few iterations. Instead, we compare the results
at the end of each step, and synchronize the parameter values so that both
implementation restart from the same point.
Also, tightened the relative error bound.