added theano flags force_device that have precedence over theano flags device when its value is not cpu. If can't use the gived device, we raise an error.

higher the relative threasold for the grad error of div and tan in float32. The problem seam to come from our way of computing the gradient. Their is not such error in float64.

make the expected numpy code of round_half_away_from_zero output the good dtype when dealing with float32.

seed the random generator to be able to reproduice the error. The buildbot use a different seed each day in debug mode and record the seed used.

fix StochasticOrder error in the buildbot. We now always do optimization in the toposort order, so they don't execute in different order.

fix TensorType.values_eq_approx. Returned False when they where equal. That fix a buildbot error. Added test for this.