Optimize NORTA setup

[matherealize]

The NORTA method implemented via simdesign_norta() requires a one time setup to compute the initial correlations for the normal distributions which are then transformed to the final marginals. This setup requires naively p(p-1) univariate optimization problems (where p is the number of desired variables), resulting in a quadratic scaling behaviour.

Can this be optimized?

• by e.g. re-using simulated normal data
• by re-using results for same marginal structures
• by optimizing the solver?

Feasible steps:

• [ ] #16