Use <cmath> for lgamma

[MichaelChirico]

This avoids any boost headaches and should be equivalent (up to machine precision).

Written by Gemini while pursuing a crash. If boost is involved in the crash, it's a bit uglier to debug vs. if std::lgamma is used. It's also consistent with nearby usages of std::log().

OTOH, AIUI boost::math::lgamma is slightly more accurate, so perhaps the choice was intentional. E.g. Gemini offers the following:

---

Where Boost is slightly more accurate:

1. Peak ULP Error:

   • boost::math::lgamma: Maintains a peak error strictly bounded under $\approx 0.6 - 0.9$ ULP across the positive real line. Boost generates its Lanczos expansion coefficients using 128-bit quad-precision arithmetic and high-degree polynomials.
   • Standard std::lgamma (glibc / libm): Peak error typically hovers around $1.0 - 2.0$ ULPs in worst-case regions.

2. Near Roots ($x \approx 1$ and $x \approx 2$):

   • Since $\ln(\Gamma(1)) = 0$ and $\ln(\Gamma(2)) = 0$, evaluating $\ln(\Gamma(x))$ near $x = 1$ or $x = 2$ suffers from catastrophic cancellation in standard polynomial approximations.
   • Boost.Math includes dedicated Taylor/Padé expansions around $x = 1$ and $x = 2$ (using log1p variants) to preserve 15–16 significant digits across the zero-crossings.

Does it matter for stochtree?

• Both implementations are accurate to 15–16 decimal digits ($\approx 10^{-16}$).
• In stochtree, lgamma(a) is called on shape parameters ($a \ge 2.0$) to compute log-marginal likelihoods. A difference of $&lt; 1$ ULP ($10^{-16}$) is negligible compared to standard floating-point addition order.