Fix inf in standard_normal

[lgarrison]

From #1, Box-Muller was using uniform floats in the half-open interval [0,1), which would cause it to generate an inf if it encountered 0 exactly. We now instead generate uniform floats in (0,1] (just for BM; uniform randoms are unchanged). This will cause the standard_normal results to change in the last few digits from what they were before.

The Box-Muller Wikipedia article says the uniform floats should be in the open interval (0,1), but this is hard to do without rejection sampling... if we are making an error, it's at most O(2^-24) for floats and less for doubles.