feat(moead): MOEA/D-DE, differential evolution in place of the crossover (Li and Zhang 2009)

[tachsin]

MOEA/D-DE (Li and Zhang 2009): differential-evolution variation for MOEA/D, for Pareto sets with complicated shapes and for constrained problems. Fan et al.'s DAS-CMOP problems (#414) need it. NSGA-II doesn't reach DAS-CMOP1, 2, 3 or 9's fronts, and neither does the paper's NSGA-II-CDP; the paper's MOEA/D-CDP with DE variation does.

API: multi::DifferentialEvolutionCrossover::new(f, cr)? goes in Moead's existing .crossover(...) slot.

• The slot takes a sealed moead::MoeadCrossover<R>, implemented for every Crossover<R> and for the DE operator on Real only. Existing code compiles unchanged, and DE on another genome is a compile error.
• As in the paper's step 2.1, a DE child whose parents came from the whole population may replace solutions anywhere, at most n_r of them. SBX keeps neighborhood-only replacement.
• Moead already compares solutions by Deb's rules (CDP), as Fan et al.'s MOEA/D-CDP does; the docs now say so.
• Python: gx.DifferentialEvolutionCrossover(f, cr, repair).

Checked in the paper: steps 2.1-2.5, eq. 6 (binomial crossover), eq. 7 (polynomial mutation), and section IV-A's settings: CR 1.0, F 0.5, η 20, p_m 1/n, T 20, δ 0.9, n_r 2. Documented differences:

• children are bred and evaluated a generation at a time;
• replacement requires strictly better;
• the ideal point uses feasible values only.

Repair, one measured deviation: an out-of-bounds gene gets a random value between the parent's gene and the bound it crossed (Repair::Bounce, the default), not a random value anywhere in its range (step 2.3's text, kept as Repair::Random). On the paper's F2, over 10 seeds:

Repair	IGD
Paper (reported mean)	0.0028
Bounce	0.0026-0.0039
Random	0.007-0.031

On DAS-CMOP (1,000 generations, 300 weight vectors, T = 30, δ = 0.2), the share of seeds reaching 99% of the front sample's hypervolume:

Problem	MOEA/D-DE	SBX MOEA/D
DAS-CMOP1	20/20	0-69% of the hypervolume
DAS-CMOP2	20/20	0-69% of the hypervolume
DAS-CMOP3	20/20	0-69% of the hypervolume
DAS-CMOP9	converges onto the front (largest distance 0.004)

DAS-CMOP9 ends at 97.3-98.0% of the hypervolume. That is above the 96.8% that even each weight vector's exact optimum would give.

Tests:

• The operator against hand-computed children, and the setting errors.
• Replacement across the whole population for DE only.
• The paper's F2 (IGD < 0.004, where SBX stays above 0.04) and ZDT1.
• A checkpoint resume.
• portable_runs entries; existing seeded Moead results are unchanged.
• Python tests.