Fuzzing special functions: sample at branch thresholds, not log-uniform

When checking incomplete gamma/beta and similar functions against mpmath, most random log-uniform samples land deep in a tail and pass. The real bugs I found sat at the algorithm's internal switch points: Boost-style igamma method 7 (x > 1000, a < x+50), and continued fractions near x = (a+1)/(a+b+2) with large a, b (jStat's 100-iteration cap). Read the dispatch code (the if/else chain picking the evaluation method), then generate x near a, near the switch point, and just either side of each numeric threshold. Run the same set through the inverse functions (gammaincinv etc.), because they inherit forward-function bugs through their Newton steps. Score inverses by backward error |F(x)-p|/min(p,1-p), so a near-zero exact quantile doesn't produce fake relative errors.