When I try to scale up my model, I find that `tf.scan` uses too much memory. Probably I have to use eager execution. How well does `HMC` and `SGMCMC` work with eager execution and tensorflow 2.0?