Add ligamma and ligamma_lower, the logarithms of tgamma(a, x) and tgamma_lower(a, x)

[NAThompson]

Logs of the non-normalised value, or lgamma plus logs of the regularized one, and a log x - x plus the logged fraction or series where that underflows. Closes #1173.

ulps against Arb (arb_hypgeom_gamma_upper for x >= a, the lower series otherwise): ligamma/ligamma_lower in blue; in orange, log(tgamma)/log(tgamma_lower), or for large a, where those overflow, lgamma(a) + lgamma_q/lgamma_p. Red is -inf or NaN. The missing orange in the last plot is lgamma_p overflowing when P underflows with a < 4x, an existing bug.

🤖 Generated with Claude Code

[NAThompson]

reporting/performance/ligamma_performance.cpp, with promote_float<false> and promote_double<false> so each type is timed in its own arithmetic. Ordinary: a log-uniform in [0.5, 30], x/a log-uniform in [0.1, 10]. Large a: a log-uniform in [1e3, 1e5], where tgamma(a, x) overflows. Median of 5 repetitions, Apple M-series, Homebrew clang 23, -O3; long double is double here.

	float	double	long double
ligamma, ordinary	66.7 ns	130 ns	127 ns
log(tgamma), ordinary	62.9 ns	125 ns	125 ns
ligamma, large a		147 ns
lgamma(a) + lgamma_q, large a		132 ns