base: fix sign confusion in log1mexp implementation (fix #17125)

author: claude (https://gitlab.haskell.org/trac-claude)

The correct threshold for log1mexp is -(log 2) with the current specification
of log1mexp. This change improves accuracy for large negative inputs.

To avoid code duplication, a small helper function is added;
it isn't the default implementation in Floating because it needs Ord.

This patch does nothing to address that the Haskell specification is
different from that in common use in other languages.