ramansep 0.4.0: peak-fitting front end

This release ships the roadmap's front end. fit_two_modes takes a raw spectrum, fits a Lorentzian to each mode window with a NumPy-only Levenberg-Marquardt solver using analytic Jacobians, and returns the peak shifts and one-sigma uncertainties that SeparationModel.invert consumes. The test suite checks the Jacobian against finite differences, exact parameter recovery on noiseless lines, statistical compatibility of the reported center uncertainty with the actual scatter on seeded noisy lines, and a full spectrum-to-inversion round trip. The fitter is deliberately Lorentzian-only and says so; a Voigt treatment for instrument-dominated lines remains future work. The README gains a peak-fitting section and the roadmap now reflects what actually shipped in each version.