Natural log of the gamma function $\ln \Gamma(x)$.

Implemented as a pure composite over engine primitives — uses the Lanczos approximation (Numerical Recipes, $g = 7$, 9 series coefficients). Relative error stays below 1.5e-15 for real x > 0 in float64; x < 1/2 goes through Euler's reflection. Autograd flows through automatically because every step is an engine op.