Fix InverseGamma.cumulativeProbability/inverseCumulativeProbability returning Gamma quantiles instead of InverseGamma quantiles

[walterxie]

Fix #149

• InverseGamma (spec) inherited cumulativeProbability/inverseCumulativeProbability from ScalarDistribution, which delegates to getApacheDistribution() — the internal Gamma(alpha, rate=beta) object used for sampling via x = 1/y. That delegation ignored the 1/y transform, so both methods returned the Gamma distribution's CDF/ICDF, not the InverseGamma's. density()/logDensity() were unaffected (they use a hand-written closed-form formula and never call getApacheDistribution()).
• Fixed by overriding both methods in InverseGamma with the correct change of variables: for Y = 1/X ~ Gamma(alpha, rate=beta), P(X ≤ x) = 1 - GammaCDF(1/x) = GammaSurvival(1/x), and inverting gives invGammaICDF(p) = 1 / GammaICDF(1-p) = 1 / GammaInverseSurvival(p). p ≤ 0 / p ≥ 1 boundaries return 0.0 / +Infinity explicitly.

Test:

• InverseGammaTest: re-enabled the previously-disabled CDF assertion in testInvGamma() against reference scipy.invgamma values (was commented out with "not implemented yet", stale even after the CDF path started silently working-but-wrong).
• Added testCumulativeProbabilityMatchesInverseGammaNotGamma: checks cumulativeProbability against reference values and the 1 - GammaCDF(1/x) formula, and asserts it no longer equals the raw Gamma CDF.
• Added testInverseCumulativeProbabilityMatchesInverseGammaNotGamma: round-trips inverseCumulativeProbability through the reference CDF values and through cumulativeProbability itself, checks the 1/GammaICDF(1-p) formula, asserts it's no longer the raw Gamma ICDF, and checks p=0/p=1 boundary behavior.
• Removed a dead commented-out Monte Carlo check in testInvGamma() that referenced a nonexistent nextInverseGamma() method (never compiled, superseded by the new deterministic tests).
• mvn test -pl beast-base -Dtest=InverseGammaTest → 3/3 pass. Full beast-base suite → 479/479 pass, 0 regressions.