Fix recurrence order in dawson's exponential-sum branch

[gaoflow]

Summary

dawson(x) returns wrong values for every |x| >= 0.2, and faddeeva(x, y)'s
imaginary part is wrong near the real axis. Both come from one bug in the
exponential-sum branch of dawson.

Root cause

src/dawson.rs, the |x| >= 0.2 branch, ports this Numerical Recipes loop:

for (i=0;i<NMAX;i++, d1+=2.0, d2-=2.0, e1*=e2)
    sum += c[i]*(e1/d1 + 1.0/(d2*e1));

The increments are in the for post-clause, so they run after each iteration
body. The port moved them to the top of the body, so the first iteration already
uses d1+2, d2-2, e1*e2 instead of the initialised d1, d2, e1. Every
term is shifted one recurrence step, so the whole branch sums the wrong series.

faddeeva inherits this: for |y| < 1e-8*|x| the imaginary part is
im_w_of_x(x) = 2/sqrt(pi) * dawson(x) (src/faddeeva.rs:280), so the
real-axis imaginary part is wrong too. The real part exp(-x^2) is unaffected.

Impact (before), oracle = mpmath sqrt(pi)/2 * exp(-x^2) * erfi(x)

x	dawson(x) before	mpmath
0.5	-1.00006	0.42444 (wrong sign)
1.0	-0.11212	0.53808
2.0	0.85660	0.30134
10.0	0.08412	0.05025

faddeeva(0.5, 0.0).im was -1.12845 vs 0.47893.

Fix

Move the three recurrence increments after the sum += accumulation, matching
the NR post-clause. No change is needed in faddeeva.rs.

Why it went unnoticed / tests

dawson_test only checked x = 0.14 (the |x| < 0.2 series branch, which was
always correct), and faddeeva_test only samples Re(z) == Im(z), so the
near-real-axis branch that calls dawson was never exercised.

Added differential tests against mpmath (agreement < 6e-8, within the documented
~2e-7 accuracy):

• dawson series branch |x| < 0.2 (still correct after the change)
• dawson exponential-sum branch: boundary x = 0.2, sign region, asymptotic
  tail to x = 10, and negatives
• odd symmetry D(-x) = -D(x) across both branches
• faddeeva imaginary part on the real axis (y = 0 and y = 1e-12*x), with
  real part exp(-x^2)

I also checked the other NR recurrence loops in the crate (betacf,
erfccheb, the Bessel routines); none share this post-clause pattern, so the
issue is localised to dawson.