Convergence tweaks for reversible MLE and sampler

[franknoe]

Affects mainly estimation: reversible tmatrix MLE and sampler:

• Added option to suppress the not-converged warning.
• Using that option and running only 100 steps in the initialization of the reversible transition matrix sampler. That's sufficient for the sampler because the MLE is just a starting point anyway. In cases where the reversible MLE doesn't converge well - and that seems to happen often for the fractional counts employed in Bayesian methods - we would otherwise optimize for 1M steps, just to throw that result away in the first sampling step. This change makes a huge performance difference for the Bayesian HMM, where a transition matrix sampling is run for every iteration of the algorithm (usually 100 times or more).