Preliminary benchmark: incremental and adaptive parallel warm-up

[palmer]

Congratulations!

@Funko_Unko

I believe the main performance gain is due to a custom warm-up procedure which not only generally performs much better than Stan’s regular warm-up, but crucially allows me to continuously adapt the “centeredness” of the parametrization, which leads to very efficient sampling. Sadly, this procedure is not yet ready for public consumption and also currently I don’t have the time to make it so.

Any plan for making it “ready for public comsumption”? Or should we settle for the current warmup for a while?