Fix assigned C=C planarity on the default inference path

[ahmedtaha100]

@jandom @jnwei Addresses the flagless reproduction in issue #136.

What changed

• Detect RDKit-assigned E/Z, non-aromatic sp2 C=C bonds and extract their heavy-atom dihedrals.
• Apply a flat-bottom correction late in diffusion, then project once more after rollout so the sampler does not undo it.
• Leave queries without matching bonds unchanged.

Verification

• On a local RTX 5090 using OpenBind 174k, seed 2746317213, and five samples, all 20 targeted dihedrals finished within 15.01° of the assigned cis/trans plane. The worst deviation fell from 101.13° to 15.0097°.
• Two focused tests cover the reported implicit-hydrogen alkene, the default no-op path, late correction, and final projection.
• 1,025 unique non-slow tests and 18 subtests pass. Three setup tests were excluded because their mock target performs real S3 checkpoint downloads. Ruff, mypy baseline comparison, diff checks, and wheel inspection also pass.

Credit

Builds on Peter Obi's (@peter-s-obi) ligand chemical-steering work. The analytical geometry is adapted from Boltz.