Not using `arma::tanh` anymore, so use a more intuitive name for the
option.

Turns out there is suppsed to be an accept-reject step after all...

Runs much faster and will less memory overhead.

Also, this commit removes the mvector header.

z sampler performance is still very poor...

Pass arma parameters by reference to the sampler. Instantaneous and
eliminates memory overhead of load() function.

Also, it's faster to process the data structure.

Will probably just eliminate the old Graph and remove the mvector file
next.

Reduces memory usage a little bit (~10%).

As a side effect, the checks for particular files when restarting bmDCA
needed to be redone due learning_rates_(step-1) not being saved anymore.

The way it was implemented created a problem when the last step saved to
the run log was one where a wait time was modified, causing determinism
to be lost when restarting.