Eliminate remaining randomness from estimation routines

Related to http://www.dynare.org/phpBB3/viewtopic.php?f=1&t=7243&p=21224#p21224 There still seem to be elements where we do not make sure that estimation becomes "deterministic". For example, we only reset the seed in metropolis_hastings_initialization after we have initialized the chains via

candidate = rand_multivariate_normal( transpose(xparam1), d * options_.mh_init_scale, npar);

thereby introducing an element of randomness. This seems to be particularly relevant if optimizers are used that rely on random numbers.