Finish numerics for Gaussian distributions

[axch]

• [ ] Find or implement erfinv (apparently CUDA has one but LLVM and libm do not) and plug it in for NormalDist quantile
• [ ] Implement the derivatives of erf and erfc so that Gaussian CDF can be differentiated (the formulas are easy)
• [ ] Implement a more numerically accurate approximation to the Gaussian CDF, e.g. with Cephes/TFP’s ndtr https://github.com/tensorflow/probability/blob/main/tensorflow_probability/python/internal/special_math.py