This fixes a regression from #15814 which unintentionally changed the return type of `Float32#abs` to `Float64`. Previously, it returned `Float32`.

`Math.copysign(self, 1)` falls back to the overload without restrictions which converts both arguments to `Float64` and thus returns `Float64`.

This patch fixes that by explicitly targeting the `Math.copysign(Float32, Float32)` overload for `Float32#abs`.
`Float64#abs` is not broken, but still applies an unnecessary integer-to-float conversion when we should be using a float right away.

I discovered this type mismatch while researching #13389