stable log cdfs for normal

[bob-carpenter]

Stan's log cdfs just take the log of the linear cdf, so they're prone to underflow to -infinity after the cdc underflows to 0 or rounds to 1.

The R implementation of log cdfs is more stable, e.g.,

> pnorm(-50, log=TRUE)
[1] -1254.831

We should be able to borrow an implementation from R or from SciPy.

[SteveBronder]

Here is the source code for pnorm https://github.com/wch/r-source/blob/trunk/src/nmath/pnorm.c#L90

[andrjohns]

The standard-normal lcdf already uses that algorithm I believe