Fix NaN in gammpapprox lower branch for large a

[gaoflow]

Problem

For a >= ASWITCH (100), gammp/gammq take the Gauss-Legendre quadrature
path in gammpapprox. Over the whole region 0.2a <= x <= a-1 they return
NaN, and invgammp panics for lower-tail p < 0.5.

gammp(100.0, 90.0)   // NaN   (mpmath P = 0.158220989186)
gammq(150.0, 120.0)  // NaN   (mpmath Q = 0.995436558696)
invgammp(0.1, 100.0) // panic: "Bad args in gammp"

These are the regularized incomplete gamma functions, i.e. the chi-square /
gamma / Poisson CDF, so for example a chi-square CDF with >= 200 degrees of
freedom (a = df/2 >= 100) is NaN over its entire lower half.

Root cause

gammpapprox computes the answer in log space:

let log_ans = log_max + sum.ln() + (xu - x).ln() + log_scale;

In the x <= a1 branch xu is set below x, so xu - x < 0 and
(xu - x).ln() is NaN, which propagates to both P and Q. invgammp
then feeds that NaN into gammp's assert!(x >= 0.0) (NaN >= 0.0 is
false) and panics.

The Numerical Recipes gammpapprox keeps ans = sum*(xu-x)*exp(...) as a
signed quantity (negative in this branch) and folds the sign into the return:
P = ans > 0 ? 1-ans : -ans, Q = ans >= 0 ? ans : 1+ans. The log-space
rewrite in 948acde kept only the x > a1 case and dropped the lower-branch
fold, while also taking ln of a negative width.

Fix

Take the log of the magnitude |xu - x| and restore the per-branch sign on
return:

• lower branch (x <= a1): P = ans, Q = 1 - ans
• upper branch (x > a1): P = 1 - ans, Q = ans (unchanged)

With gammpapprox correct, invgammp converges on both tails for large a
with no extra guard.

Verification

Reference values are mpmath (gammainc(a, 0, x, regularized=True)), stable at
dps 30 and 35. The new tests in tests/gammpapprox_large_a_test.rs cover:

• gammp/gammq for a in {100, 150, 200, 500} with x spanning the former
  NaN region, x = a-1, x = a, and x > a; mid-range values (0.158, 0.00456,
  0.472, 0.972) asserted at full precision, plus P + Q == 1 and finiteness.
• P(a, .) monotonic increasing across 0.2a .. a + 6*sqrt(a).
• invgammp round-trip both tails (gammp(a, invgammp(p, a)) == p) and a
  direct match against mpmath findroot values.
• the small-a gser/gcf path unchanged.

The old large-a quadrature test only exercised gammp(150, 150) (x = a, the
upper branch), so the lower branch was untested. Full suite passes; cargo fmt
clean; the change adds no new clippy warnings.

Fixes the large-a case of #4.