`log2mexp(x)` accuracy vanishes around the right edge of the domain, at `log(2)`

[nsajko]

The error in ULPs (here for log2mexp(x::Float64)) has a spike at the right endpoint of the domain:

The base of the spike:

A problem is that the doc string for some reason promises the exact current implementation, so fixing this bug requires a breaking change.

[tpapp]

Technically the 1.0 release did not go through, so I would be OK with a breaking change in the docs, especially if we can improve on accuracy in practice.

[tpapp]

@nsajko, if you are planning to make a fix, I can wait with 1.0 (no pressure).

[nsajko]

I'm not able to help, and don't know how difficult a fix might be, or if it is even possible.

[nsajko]

I tried employing RationalFunctionApproximation to construct an approximation, however that didn't pan out.