Statistics inv_cdf sync with corresponding random module normal distributions

[rhettinger]

Remove the inconsistent restriction that inv_cdf() is undefined when sigma is zero.

• Aligns the C _normal_dist_inv_cdf() function with its pure python equivalent.

• Restores the invariant that NormalDist(mu, sigma).inv_cdf(p) is equivalent to `NormalDist(0.0, 1.0)inv_cdf(p) * sigma + mi``. In general, operations on NormalDist should always match the rescaled and reentered results for the same operation on the standard normal distribution.

• Restores alignment with random.gauss(mu, sigma) and random.normalvariate(mu, sigma) both. of which are equivalent to sampling from NormalDist(mu, sigma).inv_cdf(random()). The two functions in the random module happy accept sigma=0 and give a well-defined result.

• This also lets the function gently handle a sigma getting smaller, eventually becoming zero. As sigma decrease, NormalDist(mu, sigma).inv_cdf(p) forms a tighter and tighter internal around mu and becoming exactly mu in the limit. For example, NormalDist(100, 1E-300).inv_cdf(0.3) cleanly evaluates to 100.0but withsigma=1e-500`` the function previously would raised an unexpected error.

• The only opposing idea is that inv_cdf() means to give the inverse of cdf(), but the cdf isn't defined with sigma=0. This is fine though. all supported inputs to cdf() are still invertible. There is just a relaxation of a restriction beyond the supported range which makes inv_cdf() obey other invariants and handle edge cases cleanly.

• The edit also gives a very small performance improvement.