[libc][math] Implement half precision lgamma function (#192834)
The implementation uses three distinct paths based on $|x|$:
- Small ($|x| < 0.66$): Fits a degree-12 Chebyshev for $h(t) =
\frac{\text{lgamma}(t) + \log(t)}{t}$, recovering $\text{lgamma}(t) = t
\cdot h - \log(t)$.
- Medium ($|x| \in [0.66, 3.37]$): Fits a degree-20 Chebyshev for $g(t)
= \frac{\text{lgamma}(t)}{(t-1)(t-2)}$
- Stirling ($|x| > 3.37$): Uses $(x-0.5) \cdot \log(x) - x +
\frac{\log(2\pi)}{2}$ plus 1/2/4/8-term Bernoulli corrections by
sub-range.
Special cases: NaN, +/-Inf -> +Inf, +/-0 pole, negative integer pole,
lgamma(1) = lgamma(2) = 0 exactly, overflow for large positive x.
Exhaustive tests pass in all 4 rounding modes.
-- CORE-MATH reciprocal throughput --
Ntrial = 20 ; Min = 391.348 + 77.447 clc/call; Median-Min = 68.093
clc/call; Max = 540.055 clc/call;
-- System LIBC reciprocal throughput --
Ntrial = 20 ; Min = 377.829 + 81.686 clc/call; Median-Min = 64.851
clc/call; Max = 532.888 clc/call;
f05a649741