Fix PRG bias

As @reyzin mentioned, naive approach to fitting 32 bits into requested range leads to bias. We need to retry with new entropy if value falls above largest number divisible by requested range. Also, would be nice to see how others implemented this routine of PRG.