Correct InverseGamma's mean and support bounds

[walterxie]

Summary:

Building on the earlier CDF/ICDF fix (#149), two more ScalarDistribution methods were found silently returning the internal Gamma helper's statistics instead of InverseGamma's own:

• getMean() was unoverridden, so it delegated to getApacheDistribution().getMean() — the internal Gamma(alpha, rate=beta) helper's mean (alpha/beta) — instead of InverseGamma's own mean (beta/(alpha-1)). Confirmed numerically: α=3, β=2 returned 1.5 instead of the correct 1.0.
• Fixed by overriding getMean() with beta/(alpha-1) for alpha > 1. For alpha ≤ 1 the defining integral diverges; since X > 0 always, that divergence is provably to +Infinity (never negative or indeterminate), so Double.POSITIVE_INFINITY is returned there rather than NaN — otherwise there'd be a discontinuity between alpha=1 (NaN) and values just above it (correctly huge, finite numbers approaching +Infinity).

Went further on getApacheDistribution() itself: it e if it were "the" Apache distribution backingInverseGamma, which is exactly what caused this bug (and the earlier CDF/ICDF one) — the Gamma object doesn't represent InverseGamma's own density/CDF/mean at all, it's onlling via x = 1/y. Per the method's own documentedcontract ("or null otherwise"), it now returns null, so any future ScalarDistribution method that isn't explicitly overridden hewill throw instead of silently returning a Gamma-flae thing that implicitly relied on it —getLowerBoundOfParameter()/getUpperBoundOfParameter() — is now overridden directly with the hardcoded correct support, (0, +Infinity), rather than derived from getApacheDistri
Test plan (InverseGammaTest, +3 tests):

• testGetMeanMatchesInverseGammaNotGamma: checks beta/(alpha-1) against the existing reference cases, plus alpha=2,beta=1 (mean still finite, but the boundary where variance stops of the mean as alpha → 1⁺, and +Infinity (not NaN)for alpha ∈ {1.0, 0.9, 0.5}.
• testGetApacheDistributionIsNull: checks the null rndOfParameter/getUpperBoundOfParameter overrides, and that inverseCumulativeProbability's p≤0/p≥1 boundary handling (which calls those) still works without getApacheDistribution().
• testSampleDrawsFromInverseGamma: sample() — the meoriginal bug report — had no test exercising it atall; this draws 200k samples and checks the empirical mean and empirical CDF against getMean()/reference values, with tolerances
  sized from the distribution's own finite variance soross repeated runs).

mvn test -Dtest=InverseGammaTest → 6/6 pass. Full beass, 0 regressions.