feat(optimizers): Add scipy-based FF optimizer

[ericchansen]

Summary

Add q2mm.optimizers -- a scipy-based force field optimization framework built on the clean model layer and OpenMM backend.

What changed

New: q2mm/optimizers/ package

objective.py -- ObjectiveFunction class

• Takes a ForceField, MM engine, molecules, and ReferenceData
• Computes weighted sum-of-squares residuals between MM-calculated and reference data
• Supports energy, frequency, bond length, and bond angle observables
• Multi-molecule training sets with per-observation weights
• Tracks evaluation count and score history for convergence analysis

scipy_opt.py -- ScipyOptimizer class

• Wraps scipy.optimize.minimize and scipy.optimize.least_squares
• Methods: L-BFGS-B (default), Nelder-Mead, Powell, trust-constr, least_squares (LM/TRF)
• Auto-bounds from ForceField.get_bounds() for bounded methods
• Tuned finite-difference step size (eps=1e-3) -- scipy's default ~1e-8 produces sub-precision energy changes for FF parameter magnitudes
• Returns OptimizationResult with success flag, scores, history, and human-readable summary

Modified: q2mm/models/forcefield.py

• Added get_bounds() method returning (min, max) tuples matching the param vector layout
• Chemical constraints: force constants >= 0, distances > 0, angles in [60, 180]

Testing

• 18 new integration tests covering all optimizer methods and data types
• L-BFGS-B, Nelder-Mead, least_squares all successfully recover known parameters
• Multi-parameter water test (bond k, r0, angle k, theta0) with frequency fitting
• Full suite: 113 passed, 22 skipped, 1 xfailed
• Validation harness: 4 passed, 0 failed, 2 blocked (unchanged)
• Ruff lint + format: clean

Design notes

The key insight for scipy integration with MM force fields: scipy's default finite-difference step for gradient estimation (~1e-8) is far too small for FF parameters with magnitudes of ~0.5-10. A step of 1e-3 reliably produces meaningful energy differences through the OpenMM evaluation pipeline. This is set via the eps parameter and passed as options['eps'] for minimize or diff_step for least_squares.