FIX: use cyclic pair ordering for LS couplings

[Zeyna777]

Closes #202

🐛 Bug fixes

• The decay node is brought into the DPD cyclic pair ordering before its Clebsch-Gordan factors are formulated, so the $LS$ couplings and the isobar Wigner-d of that same node agree on which decay product comes first. The inline swap suggested in LS Clebsch-Gordan factors use a different pair ordering than the isobar Wigner-d function #202 is extracted into a documented _order_decay_products helper, which leaves one call site:

  amplitude *= _formulate_clebsch_gordan_factors(
      isobar=_order_decay_products(chain.decay_node, first_id=i),
      helicities={...},
  )

• Of the two options in LS Clebsch-Gordan factors use a different pair ordering than the isobar Wigner-d function #202, this takes the one that fixes it in the builder: to_three_body_decay keeps sorting by final-state ID, so ThreeBodyDecayChain.decay_products — and therefore the dynamics builders and the serialization round-trip — stay untouched.

❗ Behavioral changes

• Affected models need a refit, not just a re-evaluation: the amplitude of every $LS$ model that populates subsystem 2 shifts by a wave-dependent sign. Models in the helicity basis, and $LS$ models without a subsystem-2 chain, are unaffected.

🖱️ Developer experience

• test_ls_amplitude_does_not_depend_on_child_order states the invariant directly, rather than through the C-symmetry observable of LS Clebsch-Gordan factors use a different pair ordering than the isobar Wigner-d function #202: which of the two children an IsobarNode happens to store first is an implementation detail of whoever built the ThreeBodyDecay, so reversing it may not change the amplitude. The test builds each subsystem amplitude twice, once from the decay as to_three_body_decay produces it and once with every decay node's children reversed, and asserts that the two expressions are symbolically equal.

  Without the fix it fails on the subsystem whose ordering flips, with a relative sign:

  A: -(sin(theta_12/2)/4 - 3*sin(3*theta_12/2)/4)*H[N(1700)+, 2, 1/2]*H[N(1700)+, 1, 1]/4
  B: +(sin(theta_12/2)/4 - 3*sin(3*theta_12/2)/4)*H[N(1700)+, 2, 1/2]*H[N(1700)+, 1, 1]/4
  

• No existing test or notebook covered this: the three docs/comparison notebooks that assert DPD == AmpForm all use min_ls=True, docs/xib2pkk.ipynb uses the $LS$ basis but only asserts non-NaN, and docs/lc2pkpi.ipynb uses min_ls=(False, True), whose decay node is in the helicity basis and therefore unaffected. The whole suite passes unchanged either way, which is why this went unnoticed.

[redeboer]

Thanks! 🐉
Verified on ComPWA/polarimetry@cf68fe0 (ComPWA/polarimetry#444), with decay rates for the LS model asserted to be unaffected 🎉