docs: Improve tutorial, fix benchmarks, add molecule constructors

[ericchansen]

Summary

Improve tutorial docs from live review, fix flawed performance benchmarks, and enhance API reference.

Changes

Tutorial (docs/tutorial.md)

• Add collapsible molecule constructor examples (Gaussian GaussLog, QCElemental, raw arrays)
• Connect Step 1 QM outputs to Step 2 molecule building with intro text
• Combine two redundant bond_tolerance info boxes into one comprehensive explanation
• Add Hessian eigenmatrix training data warning (core legacy capability not yet ported -- see feat: Port Hessian eigenmatrix training data with TS coordinate inversion #64)
• Replace hardcoded ReferenceData values with loop-based extraction from mol.bonds/mol.angles
• Add note about future ReferenceData.from_molecule() automation (see feat: Auto-populate ReferenceData from QM outputs and molecules #63)

Performance benchmarks (docs/performance.md)

• Fix flawed data: L-BFGS-B and Powell previously showed ~0 initial scores because they were run on already-optimized parameters instead of fresh starting points
• Regenerate all numbers from scripts/generate_benchmarks.py with identical starting params (initial score 42.9 for all methods)
• Add L-BFGS-B convergence warning (final score 0.93 vs 0.000 for Nelder-Mead/Powell)
• Update throughput: OpenMM ~190 eval/s (was ~10), Tinker ~1.3 eval/s (was ~6)

API reference (docs/api.md)

• Add Q2MMMolecule constructors table (from_xyz, from_structure, from_qcel, raw)
• Add Gaussian log file example
• Cross-link to tutorial for bond_tolerance details

New script

• scripts/generate_benchmarks.py -- reproducible benchmark generation so performance data is never hand-written again

Related Issues

• feat: Auto-populate ReferenceData from QM outputs and molecules #63 -- Auto-populate ReferenceData from QM outputs
• feat: Port Hessian eigenmatrix training data with TS coordinate inversion #64 -- Port Hessian eigenmatrix training data
• feat: Export optimised ForceField to OpenMM XML format #65 -- Export ForceField to OpenMM XML

Testing

• mkdocs build --strict passes clean
• ruff check . passes
• Benchmark script verified: all methods start from identical params (score 42.9)