accuracy of floating point numbers in python