Fix DivisionByZeroError in the hypergeometric family for alternating series

[gaoflow]

The shared generalizedHypergeometric() convergence loop tests

while (\abs($product / $sum) > $tol && $iteration < self::MAX_ITERATIONS);

which divides by the running partial sum $sum. For any alternating series the partial sum passes through exactly 0 on an early iteration, so this raises DivisionByZeroError on finite, well-defined inputs. It reaches the public confluentHypergeometric (₁F₁) and hypergeometric (₂F₁) entry points:

Special::confluentHypergeometric(1, 1, -1);  // ₁F₁(1;1;-1) = e⁻¹ ≈ 0.36788, throws
Special::confluentHypergeometric(2, 2, -1);  // throws
Special::hypergeometric(2, 2, 2, -0.5);      // ₂F₁(2,2,2,z) = (1-z)⁻² = 0.44444, throws

₁F₁(a;a;z) ≡ eᶻ is a textbook identity, so the first case unambiguously should return e⁻¹.

A grid sweep of ₁F₁ (a ∈ {-0.5, 0.5, 1, 1.5, 2, 3}, b ∈ {0.5, 1, 1.5, 2, 3}, z ∈ -3..5) and ₂F₁ (|z| < 1) found 15 crashing inputs, all the same root cause, all on finite values.

Fix

Compare the term magnitude to $tol * \abs($sum) (multiply) instead of \abs($product / $sum) (divide). Multiplying can never divide by zero, and for a non-zero $sum it is algebraically identical to the old test, so every already-correct result is unchanged. When the partial sum is momentarily zero the loop simply keeps iterating. The MAX_ITERATIONS guard is retained.

The sibling lowerIncompleteGammaSeries() in the same file already guards the identical divide-by-accumulator (\abs($sum) > 0 && \abs($term / $sum) < $tol); this brings the hypergeometric loop in line with it.

No regression

I compared the fixed output against the unpatched output across the full grid: all 500+ currently-working inputs are bit-for-bit identical (as expected, since the convergence test is only algebraically rearranged), and the 15 formerly-crashing inputs now match mpmath hyp1f1/hyp2f1 (dps 30 and 35 in agreement) to the series' ~1e-8 tolerance.

Tests

Added data-provider regression tests across ₁F₁ and ₂F₁ for the zero-crossing inputs, plus the ₁F₁(a;a;z) = eᶻ axiom over the sign of z. Expected values are from mpmath (dps 45 and 50 agree). Full suite is green (phpunit, phpcs, phpstan).

[gaoflow]

The red jobs aren't from this change: all 53 failures are ReflectionException from ReflectionMethod invocations on private Piecewise/Polynomial/Finance/chebyshevEval methods in files this PR doesn't touch, and #491 (a disjoint diff) fails with the identical list. The same tests pass locally with the repo's pinned PHPUnit 8.5.53, so it looks like a CI runner/PHP-build issue.