Whitelist a log1p failure in newlib
This might be an actaul newlib bug: it's not providing a range error when taking log1p of -1e80, but it appears that this triggered a glibc bug from a while ago that may have not been fixed in newlib. I'm not sure why the bug is just manifesting now (the newlib sources haven't changed for a long time), but the glibc version has some macros that force the evaluation of particular floating-point expressions (one of which produces the range error) that don't exist in newlib. I'm OK whitelisting the error for the next release, despite it technically being a regression. log1p isn't a standard function, and I doubt embedded users are going to notice. That said, we should still fix it.