Add EnzymeGammaExt with analytic derivative rule for Gamma.gamma

[oschulz]

Claude Opus (max effort) says:

HypergeometricFunctions v0.3.29 switched its gamma import from
SpecialFunctions to the Gamma package. Unlike SpecialFunctions.gamma
(a ccall to tgamma), Gamma.gamma is a pure-Julia Cephes implementation
with internal while loops. When it is inlined into a larger reverse-mode
gradient -- e.g. logpdf(truncated(Beta(a,b),lo,hi), x) differentiated
w.r.t. the shape parameters, which routes through the Beta CDF and the
2F1 hypergeometric function -- Enzyme cannot rematerialize the loop and
fails in libEnzyme with an internal "ficticious phi" recompute error.
This is what breaks the "truncated Beta" case of the Distributions
integration test on Julia 1.10.

Add a Gamma package extension supplying the analytic derivative
d/dx gamma(x) = gamma(x) * digamma(x), so Enzyme differentiates the
closed form instead of descending into the loops. This mirrors the
existing special-function rules (e.g. _logabsgamma, beta_inc).

Add test/ext/gamma.jl covering the scalar derivative across every branch
of the implementation and the previously-failing in-context 2F1 case.

[oschulz]

Should I reduce/remove the AI code comments?