Fix odd-N mating and add an opt-in two-shuffle pool

[jkbennitt]

Summary

Odd populations skipped parents[N-1] in SBX because the pair cursor wrapped to 0 before that index was read. Das–Dennis sizes such as M=2, p=12 (N=13) and M=3, p=12 (N=91) hit this every generation. The leftover parent is now paired with parent 0, then the cursor restarts. Even N still walks (0,1), (2,3), … and wraps to (0,1) with no extra random draw, so the published ZDT1 (52) and DTLZ2 (92) pair sequences stay the same.

Two opt-in modes follow the existing TournamentMode style. Defaults are unchanged.

• MatingPoolMode.TwoShuffledPasses: two shuffled passes of consecutive tournaments (Seada & Deb Algorithm 1 and pymoo 0.6.2 TournamentSelection with pressure 2). When N parents are drawn from N members, each index is a contestant twice. The default pool remains N independent tournaments with replacement, which is the published ZDT1 Wilcoxon protocol.
• TournamentMode.Algorithm2: a same-niche perpendicular-distance tie keeps the second parent. PymooCompatible still coin-flips. RankNicheDistance is unchanged. Either comparator works with either pool.

No package version bump. No new IGD or Wilcoxon numbers.

Checklist

• dotnet test Unsga3.slnx -c Release passes locally
• CHANGELOG.md updated under Unreleased (if user-visible)
• Public API changes documented (XML docs / README)
• If survival / normalization / metrics changed: noted oracle impact (or re-ran multi-seed)

Survival, normalization, duplicate keys, p_c, and the published tables are untouched. Even-N mating order is locked by tests, so the published ZDT1 / DTLZ2 protocols are not remeasured.

Test plan

• dotnet test Unsga3.slnx -c Release — 87 passed, 0 failed.
• Odd N: AdvanceMatingPair and a recording crossover prove parents[N-1] is an SBX parent for N=13 (and the cursor for N=3 and N=91).
• Even N: consecutive pairs for 4, 52, and 92, including the wrap back to (0,1). A recording crossover checks the N=4 cycle does not use the odd wrap.
• Two-shuffle: for N in {4, 13, 52, 91} and 40 seeds, every index is a contestant twice and every contest is two distinct members. SelectParents follows that contest order.
• Default pool: independent Next / NextExcept draws still match the previous sequence on a ranked population.
• Algorithm2 keeps the second parent on a same-niche distance tie, still prefers the shorter distance, and still coin-flips an equal constraint violation. PymooCompatible coin-flip tests are unchanged.
• One-generation ZDT1 run with Das–Dennis (2, 2) (N=3), TwoShuffledPasses, and Algorithm2 returns a full population.

What stayed unchanged

• Default tournament RankNicheDistance
• Default mating pool (independent tournaments with replacement)
• PymooCompatible distance-tie coin flip
• p_c = 1.0
• Duplicate G12 keys
• Collapsed nadir = ideal + 1
• PrepareForSelection re-normalization
• Published IGD / Wilcoxon tables