Derivation of Probabilistic Diffusion Model

Question

In the diffusion model tutorial (paper), the consistency term in equation 21 is reformulated into equation 22 by reversing the probability process from $$q(x_(t)|x_(t-1))$$ to $$q(x_(t-1)|x_t, x_0)$$. I understand how the individual term is reformulated but it is not very clear with the expectation operation and the KL divergence. In short, how equation (21) is reformulated to equation (22)?

I tried to expand equation 21 into the integral forms and substitude the probability reverse into the integral form, but still could not get the result in equation 22.