Reduce Ewald summation work with pair-specific image bounds

[antoine-levitt]

Cf #1387. I haven't checked any of it, but the basic idea appears right.

AI slop follows.

Ewald setup evaluates every atom pair over a global box of periodic images, including many whose screened interaction is negligible. Bound the images separately for each pair and apply the real- and reciprocal-space spherical cutoffs before evaluating the expensive terms. This addresses the bottleneck discussed in #1387.

The real-space loop now updates Cartesian separations incrementally and converts accumulated Cartesian forces to reduced coordinates once per atom. It also uses the analytical radial derivative and accumulates energy per atom to reduce roundoff. The default η and full ordered pair sum are unchanged. Complex phonon displacements are included in the image bounds; their derivatives still use the complex-analytic distance.

This also fixes the ordinary energy_forces_ewald wrapper silently discarding the supplied η.

CPU timings

216 Al ions in a 6×6×6 FCC primitive supercell, lattice parameter 4.05 Å, Julia 1.12.4, Alder Lake CPU. This is a representative system; #1387 does not provide its complete benchmark input.

Implementation	Float64 Ewald call	Six strain directions, Ewald only
Master	2.24 s	20.55 s
#1387 at 665d824	0.83 s	7.61 s
This PR	0.080 s	1.09 s

Compilation is excluded. Float64 timings are minima of three warmed calls. The strain timings are the measured sum of six one-partial Dual calls, matching the Ewald work in compute_stresses_cart; they do not include the rest of stress evaluation. Ewald runs on the CPU even in GPU calculations. No GPU benchmark was run.

benchmark/ewald.jl includes the reproducer and instructions for comparing the implementations.

Validation

• All 19 assertions in the three energy_forces_ewald test items pass: existing reference energies, η independence and forwarding, periodic shifts, forces, strain derivatives, and a nonzero-q complex phonon response checked against finite differences.
• Independent comparisons with the original implementation on a skew three-ion cell give force and first/second phonon-response errors below 4×10⁻¹⁴.
• The 216-atom energy agrees with 216 times the primitive-cell energy within 1.4×10⁻¹² Ha. Perturbed Al216 forces agree with the original to 4.6×10⁻¹⁴ relative error.

Targeted test command in the persistent test environment:

TestItemRunner.run_tests("test/";
    filter=ti -> startswith(ti.name, "energy_forces_ewald"))

[abussy]

I've tested this code on my 216 Al atoms system, the Ewald term instantiation takes ~900 ms vs 23 seconds initially, and 9.5 seconds with #1387.

The AI code looks clean and lean enough too. I'm happy to have this solution over #1387, but some clean up is needed.

[antoine-levitt]

I'm happy to have this solution over #1387, but some clean up is needed.

Can you do it? Otherwise I can take a look but it might take a while