[WIP] RNG test fixes and improvements

[vinaydes]

• Fixes the failures of RNG tests and issues mentioned here.
• Reference standard deviation calculations: https://gist.github.com/vinaydes/cee04f50ff7e3365759603d39b7e079b
• Additionally fixes issue of Rng throwing NaNs and Infs
• Need to add parameter validation for each distribution

@MatthiasKohl @teju85 Please take a look.

[vinaydes]

I'll take a look at failing test Rng.MeanError.

[vinaydes]

Summary of latest 3 commits:

1. Make uniform generation same for both PCG and cuRAND - [0.0, 1.0)
2. Change seed of GEMM unit test since it was generating too small floating point values to compare correctly
3. Fix Inf/NaN occurrences by adding 0.0 checks at multiple places

[MatthiasKohl]

I went over all distributions again (not just the ones you changed). Things I noticed:

1. For Bernoulli / ScaledBernoulli, I'm almost sure the calculation is simply wrong: the probability parameter usually represents chance of success, but with res > params.prob this would mean that for a probability of 40% we get success 60% of the time. So I guess here probability represents non-success. But, if params.prob is 0, we can still get non-success if res == 0. So I'm almost certain that it should actually be res < params.prob: For params.prob == 1, we always get success, for params.prob == 0, we always get non-success and everything in-between should work fine.
2. In Laplace, there is something similar to the above, but it is correct in the code: I would still add a comment that since res is in [1, 2^24 - 1] / 2^24, we will have 2^23 numbers in the first branch and 2^23 - 1 numbers in the second one. The special value 2^23 / 2^24 == oneHalf would give the same result anyway in both branches, so it doesn't matter where it goes. This is probably not obvious at first sight for someone going over this code.

Apart from that, everything else looks good to me.

[MatthiasKohl]

Changes LGTM from my side

[cjnolet]

LGTM

[cjnolet]

@gpucibot merge