AMDGPU: Correct constants used in fast math log expansion

The division between float constants was done with less
precision. Performing the divide in double and truncating to float
provides the same value as used in the library fast math expansion.