Evaluate only desired branch in log1mexp_numpy

[ricardoV94]

This PR changes log1mexp_numpy in the math module to only apply one of the two operations to the inputs. This avoids the non-important but common RuntimeWarning during NUTS sampling when logdiffexp_numpy is called in the stats method:

RuntimeWarning: divide by zero encountered in log1p
return np.where(x < 0.6931471805599453, np.log(-np.expm1(-x)), np.log1p(-np.exp(-x)))

The RuntimeWarning occurred because both operations were applied to x before selecting the desired (more numerical stable) output with np.where.

I added a test that fails in Master but passes in this PR.

[ricardoV94]

Maybe this shouldn't change x in place...

[ricardoV94]

Also, if the new unittests feel like too much, I am happy to remove them.

[michaelosthege]

Looks good now. The tests are fine too.

I think you should add a line to the release notes - I also added this to a "vNext (3.11.1)" milestone.

[ricardoV94]

Added. I also had to create the new "vNext (on deck)" section.