Add fast_compile tag so the optimization to remove the Op always runs. Use g_out.zeros_like instead of tensor.zeros_like so the op also works for sparse variables.