use von_mises_lpdf directly instead of von_mises2_lpdf after next stan update

[venpopov]

Currently the von_mises() family is not vectorized if kappa is predicted, because the von_mises_lpdf is numerically unstable for large kappas, and you use the normal approximation for kappa > 100, which requires an if else check.

I have submitted a PR in stan math library that improves the stability of the von_mises_lpdf and which will make the von_mises_lpdf2 unneccessary, as I describe in the issue here.

Once that is accepted and released, you could simplify the von mises family, which my tests show it speeds it up, and also avoids scale=0 warning when 1/sqrt(kappa) passed to normal_lpdf is 0. Some tests described here at the end.

It's probably going to be a while until the next release and also that will work just for the cmdstar backend until rstan catches up, but I'm nothing it here for when the time is right.