[Tests] Fix precision constraints for FP math operations

Some floating point operations rely on approximation. The problem is to test such functions against standard version with automatically generated data inputs.

What should be done is: 1. design a code pattern for calculating 'ULP distance' from the reference value, 2. create a test fail/pass threshold value expressed in this ULP distance, 3. report the ulp precision (for selected tests only) together with test result.