Improving computational efficiency of hyper-parameter tuning

[boyuren158]

Currently the program runs through a grid of hyper-parameter values and picks the one minimizing the covariate balance. This approach is very slow and might spend too much time on "unlikely" configurations. An alternative approach of hyper-parameter tuning can be achieved by using covariate balance as a pseudo-likelihood and combining a prior distribution of the hyper-parameters with it to enable a Bayesian inference.