Address floating point issues

[lumip]

It is known that sampling ideal DP noise using floating point is vulnerable to certain attacks and solutions have been proposed (e.g., the discrete Gaussian mechanism, interval refining sampling and others). We need to implement one of these.