Test failure is a timeout

This is a good test in theory, but I have a few issues with it in practice. The first is that the default rtol for ≈ is sqrt(eps(T)) which is much bigger than the 2eps(T) we probably want for scalar functions. The second issue is that since this only tests a single value, it is reasonably likely that a mostly good log implementation will get this value correct by chance. A more thorough test would be to take 1000 equally spaced subnormals, and perform a test like isapprox(T(log(widen(x))), log(x), rtol=2*eps(x)

That's a good point, I'll add a more thorough test of various values and tighten the bounds

Also while I'm making suggestions, any chance you could add log2 and log10 tests here? It will be nice (especially since #39556 will hopefully be changing the implementation of log2 and log10)

Alright, done! Look good to you, @oscardssmith ?