Implement lazy evaluation for BayesMBAR to avoid unnecessary computations

[xiki-tempula]

Summary

• Refactor BayesMBAR to use lazy evaluation, deferring expensive computations until the relevant properties are accessed
• When only F_mode is needed, MCMC sampling is now completely skipped, significantly improving performance

Details

Before: All computations (mode + MCMC sampling) happened in __init__, even if only F_mode was needed.

After: Computations are deferred and only triggered when the corresponding property is accessed:

Property	Uniform Prior	Normal Prior
F_mode	Mode only	Mode + samples (for hyperparams) + posterior mode
F_mean, F_cov, F_std, F_samples	Mode + sampling	Full posterior sampling

This is particularly beneficial for users who only need point estimates (F_mode) and don't require uncertainty quantification, as they now avoid the expensive MCMC warmup (500 steps) and sampling (1000 samples).