feat: kekulize aromatic ions, delocalized ring anions, and connected organometallics

[fbaensch-beilstein]

Stacked on #239 (base branch extract-aromaticity-utils, which adds the aromaticity.{c,h} module this builds on). GitHub will retarget this to dev automatically once #239 merges. Supersedes #234; refs #154, #82.

What this does

Aromatic systems that previously failed with Error -9986 (Cannot process aromatic bonds) now kekulize into valid InChIs. InChI has no "aromatic" bond order — type-4 bonds are kekulized into localized single/double bonds by a balanced-network search that needs a perfect double-bond matching. Odd rings whose charge/radical center contributes an empty orbital, lone pair, or single electron (not a double bond) have no such matching and fail.

The fix adds a Hückel-informed relaxation on the failure path only, then rebuilds the network from the original flexible bonds and retries once:

1. A self-contained classifier in aromaticity.{c,h} — is_aromatic_electron_source (a ring atom that can't take a double bond: ring-coordination 2 excluding metal bonds, spare valence, charged or doublet-radical) and relax_aromatic_electron_sources (moves one valence unit to an implicit H). Fully unit-tested; no BN_STRUCT dependency.
2. In mark_alt_bonds_and_taut_groups (ichi_bns.c), on BNS_ALTBOND_ERR: restore a pre-search snapshot (flexible type-4 BOND_ALTERN bonds + original valences), relax the electron sources, tear down and rebuild the network from a fresh AllocateAndInitBnStruct (so vertex capacities re-derive from the relaxed atoms), and retry the conversion once. Restoring the flexible bonds lets the search itself find the matching — so charge position stops mattering and metal-bonded rings resolve in place.

Because it runs only after the first kekulization already errored, every structure that kekulized before is byte-identical.

Now processes correctly

Structure	Result
Cyclopropenyl cation (C₃H₃⁺)	InChI=1S/C3H3/c1-2-3-1/h1-3H/q+1
Cyclopentadienyl anion (C₅H₅⁻)	InChI=1S/C5H5/c1-2-4-5-3-1/h1-5H/q-1
Tropylium cation (C₇H₇⁺)	InChI=1S/C7H7/c1-2-4-6-7-5-3-1/h1-7H/q+1
Methyl-cyclopentadienyl anion — all 5 ring −1 placements	one identical InChI=1S/C6H7/c1-6-4-2-3-5-6/h2-5H,1H3/q-1 (charge-placement invariance)
Ferrocene (connected, -MolecularInorganics)	InChI=1B/C10H10Fe/…/q-2
Cobaltocene anion (connected, -MolecularInorganics)	InChI=1B/C10H10Co/…/q-1
Pyridinium (control)	unchanged — 3-coordinate N⁺ correctly not relaxed

Deferred (strict xfail, flip loudly when fixed)

• Neutral cyclopentadienyl radical — fails upstream in the pre-existing FIX_AROM_RADICAL path, before this hook.
• Default-mode (disconnected) organometallics — a distinct failure in the metal-disconnection path; the connected -MolecularInorganics path is the one addressed here.
• Neutral-drawn metallocene rings — no ring charge to relax (chemically under-specified; the charge-explicit drawing works).

Verification

• Clean full build (CLI + libinchi).
• C++ unit tests: 17/17 suites (22 test_aromaticity cases).
• Executable pytest: test_aromatic_ions.py 7 passed / 2 xfailed; full suite 19 passed / 3 skipped / 9 xfailed, no failures.
• Regression suite (config_ci, 4190 structures): 0 diffs.
• Multithreading: clean.

Guarantee: no regression on any structure that already produced an InChI; new InChIs appear only for inputs that previously errored with -9986. The one-shot retry either produces a valid kekulization or preserves the original error.

Relationship to #234

Supersedes #234 (fix/aromatic-ion-valence), whose charge-center relaxation was the starting point; its test_aromatic_ions.py cases are carried over here. #234 can be closed in favor of this.