LogNormal Variance loses precision for small σ and returns 0 below σ≈1e-8 #84

What's wrong

Essentials.DistributionProviders.LogNormal/LogNormalDistributionProvider.cs:59-66:

double logVariance = LogStandardDeviation * LogStandardDeviation;
return (Math.Exp(logVariance) - 1.0) * Math.Exp((2.0 * LogMean) + logVariance);

Math.Exp(σ²) - 1.0 cancels catastrophically for small σ²: relative error grows like ε/σ², and once σ² < ~1.1e-16 the result is exactly 0.

Measured on net10.0 with the expression copied verbatim (expected = ExpM1(σ²)·exp(2μ+σ²)):

Independent of logMean, so LogNormal(5, 1e-8) reports Variance == 0 and StandardDeviation == 0 while Mean ≈ 148.

Why it matters

Same defect class as the recently fixed #49 (Triangular variance cancellation) and #52 (Exponential CDF near zero). Users modelling tight multiplicative noise (log-scale σ of 1e-6…1e-9) get a variance wrong in the leading digits, or zero. The cure is already in the project: Shared/SpecialFunctions.cs:149 has a cancellation-free ExpM1, and it is already linked into the LogNormal project.

Suggested fix / acceptance criteria

• Variance = SpecialFunctions.ExpM1(logVariance) * Math.Exp((2.0 * LogMean) + logVariance).
• Tests: (0, 1e-8).Variance ≈ 1e-16 (rel. 1e-12); (0, 1e-6).Variance ≈ 1.0000000000015e-12; existing LogNormal(0,1) value unchanged.