Widen PolyLog(s, z) to a non-integer or complex order; Part of #340

[enumeratio]

Part of #340. Depends on #356 (LerchPhi): this branch is on top of it.

PolyLog(s, z) now evaluates at non-integer and complex order through Liₛ(z) = z·Φ(z, s, 1), reusing the LerchPhi kernel rather than a second implementation. PolyLog(s, 1) and PolyLog(s, −1) reduce exactly to Zeta(s) and (2^(1−s) − 1)·ζ(s) for any order. Integer orders are unchanged. Compiles to JavaScript, GLSL and WGSL for real operands.

Numeric policy as for LerchPhi: it stays unevaluated where it can't be accurate. That's past |z| = 1 where LerchPhi declines (#353), and three measured zones near the branch point z = 1, near a positive integer order on the rim, and for real z < 0 with very negative order. Checked against mpmath polylog over s ∈ [−4, 6] (real and complex) and z inside, on and outside the unit disk, on both sides of the cut: no answered value off by more than 6e-13.

[arnog]

Thanks, merged (after #356), with a review pass on your branch as 145cdaed. The main corrections:

• The compiled JavaScript and GPU cases sent every order through the widening kernel, so compiled PolyLog(3, -2) was NaN and PolyLog(0, 1) was $−0.5$ instead of the pole. A dispatcher now handles $z = 0$, the orders $1, 0, −1$ , integer orders $≥ 2$ and the rational forms for $−2..−12$ first, for both cases.
• The near-integer guard looked only at $Re(s): s = 5 + 1e-6i$ on the rim returned values off by up to $3e−4$. The distance is now measured in the complex plane.
• The whole real axis $z &lt; −1$ declined for every non-integer order although the value is real; Jonquière's inversion answers it where it measures accurate (PolyLog(1.5, -3) = −1.679089730504828). What remains is tracked in ROADMAP.
• The $z = 1$ guard declined $0.98 &lt; |z| &lt; 1.02$; measured, the kernel is accurate to $1e−12$ down to $|z − 1| = 1e−3$, which is the new radius. The series guard no longer declines a positive real $z$ with a negative order.
• Negative integer orders $−2..−12$ use their rational closed form (PolyLog(-2, 1/2) is the exact 6); the type handler claims real only for $z$ proven $≤ 1$ .

Two limits of the current hurwitzZetaComplex (a negative order with a complex base point; the machine zeta near a small negative order) are recorded in ROADMAP; they bound the inversion.