The quadrature tests lower the the error substantially for beta Bernoulli and Normal.
Fixed two bugs in normal: Welford's algorithm wasn't accurately calculating the mean and variance (for incorporate / unincorporated) and replaced a difference of squares with a product (for numerical stability)
Changed adapter to test that it produces the same result as the inner distribution (since that's what we actually care about).