Implement LerchPhi; Part of #340

[enumeratio]

Part of #340

Adds LerchPhi(z, s, a) = Σ zᵏ (k+a)^(−s), Wolfram's LerchPhi. LerchPhi(1, s, a) reduces to HurwitzZeta(s, a), and a non-positive integer a is a pole when Re(s) > 0, as for HurwitzZeta. Inside the unit disk it's direct summation, on the real rim a van Wijngaarden transform, and past |z| = 1 the Hermite-type integral through the complex incomplete gamma. Machine precision. Compiles to JavaScript, GLSL and WGSL for real operands.

It never returns a value it can't vouch for. Past the unit disk it stays unevaluated where the incomplete gamma argument −a·log z has Re < 0 and modulus above 2.5: measured against mpmath, incompleteGammaUpperComplex loses digits there (#353, which is wider than that issue first described). Fixing #353 would widen LerchPhi with it. Checked against mpmath lerchphi on a random 1,500-point sweep across the disk, the rim and past it, with no answered value off by more than 4e-13.

[arnog]

Thanks, merged. Before merging I ran a review pass (every value checked against mpmath) and pushed the results onto your branch as f20b2faa, on top of a merge of main. The commit message lists each change; the main ones:

• The direct series and the Euler transform returned unconverged partial sums (one small increment counted as convergence): LerchPhi(0.5, -2, -9.000000001) dropped a tail of $0.0117$ and LerchPhi(-0.99, -12, 1) gave $2.25e12$ where the value is $−13.85$. Terms at non-positive base points are summed first, convergence needs three consecutive small and decreasing increments with a tail bound, and the kernel declines instead of returning a wrong value.
• The base-point shift never terminated for a large negative $a$; it now declines past $1e6$ steps.
• The type handler claimed real for $z$ on the cut and at the pole; the compiled wrapper skipped the pole checks; the GPU kernels returned truncated sums and rejected $z = −1$; a float operand gave exact results. All fixed, with an f32 simulation of the shader against mpmath.
• Accuracy is stated as measured (about $1e−11$) rather than machine precision; the ROADMAP gap is GPU-only, and the real $z &lt; −1$ decline is tied to Gamma(s, x) near the negative real axis drops its branch term #353.

One thing coming your way: the #353 fix (a new incomplete-gamma kernel for Re(x) < 0) lands next. incompleteGammaRelativeError estimates the old kernel's error per branch, so I'll adapt that estimate when #353 lands; expect the decline set past $|z| = 1$ to shrink.