Feature/common random

[adolgert]

This adds one way to implement Common Random Numbers (CRN). This is a method for reducing the number of draws required in order to get good statistical power on an estimate of the effect of changes to parameters in a simulation. The idea is to run a simulation once, recording its randomness, then for each modification of the parameters, run the simulation with the same underlying randomness. This will destroy any measurements of the dispersion of trajectories, but it will tell you centrality measures with many fewer draws.

This implementation records the state of the random number generator before each and every enabling of a distribution. That would be crazy to do with a Mersenne twister random number generator because its state is large, but a Xoshiro generator has 32 bytes. This is still large, but is manageable. The advantage to this, over recording draws from uniform numbers, is that it lets the sampler use the fastest available methods to sample distributions. Disadvantages are that it needs more memory.