Fix flaky ChainRules AD tests by keeping finite differences off domain boundaries

[devmotion]

Fixes #176

Problem

The AD tests in test/chainrules.jl intermittently fail in CI with DomainErrors from loggamma/digamma/log, e.g. this run. The failures have appeared on master and several PRs and are not caused by any recent change.

The analytic rules in ext/StatsFunsChainRulesCoreExt.jl are correct. The flakiness comes from the numerical reference: test_frule/test_rrule perturb the inputs with finite differences, and with unseeded exp(randn())/logistic(randn()) draws an argument occasionally lands close to a domain boundary, so the stencil crosses it and the primal throws (e.g. logbeta(α, β) with β < 0).

Fix

Keep every finite-difference stencil inside the domain by anchoring the test points well away from their boundaries and relying on ChainRulesTestUtils' defaults — no custom fdm or tangents needed.

test_frule/test_rrule evaluate the reference at input ± reach, where reach = max_range · |tangent|. With CRTU's defaults the reach is bounded:

• max_range = 1e-2 (CRTU's _fdm = central_fdm(5, 1; max_range = 1e-2));
• Float64 tangents are drawn from -9:0.01:9, so |tangent| ≤ 9 (they are not unbounded randn — that applies only to the generic/BigFloat methods, not to the Float64 arguments here).

Hence reach ≤ 1e-2 · 9 = 0.09. Drawing every differentiated argument ≥ 0.1 from its boundary (pos() = 0.1 + randexp(), unit01() = 0.1 + 0.8rand()) makes a boundary crossing impossible by construction (0.09 < 0.1), for any draw or seed. Each log-pdf domain is an independent box constraint (α>0, β>0, x∈(0,1), ν>0, λ>0, …), so per-coordinate anchoring suffices under simultaneous perturbation. The wider margin also keeps the points better-conditioned than the singular regions, so the finite-difference reference stays accurate at the default rtol. A fixed seed is added for reproducibility, and integer arguments resolve to NoTangent() automatically.

This margin depends on CRTU's defaults; a comment in the test records that if max_range or the tangent range grows, the 0.1 anchor must grow with it. FiniteDifferences is not added as a dependency.

Verification

• 0 failures over 2500 seeds × 7 distributions (frule + rrule) at the default tolerance (240000/240000 checks pass).
• Full Pkg.test() passes locally (chainrules | 98 98); the QA testset (Aqua/ExplicitImports/JET) stays green after dropping the FiniteDifferences test dependency.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 72.39%. Comparing base (f8d7f5b) to head (709893b).
⚠️ Report is 2 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #215   +/-   ##
=======================================
  Coverage   72.39%   72.39%           
=======================================
  Files          22       22           
  Lines        1007     1007           
=======================================
  Hits          729      729           
  Misses        278      278           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.