overdetermined multimode inversion with model checking

New module multimode: the (strain, density) inversion for any number of Raman modes.

• MultiModeModel.invert: per-pixel generalized-least-squares (Gauss-Markov) inversion, weighting each mode by its shift uncertainty; vectorized over whole maps.
• • Per-pixel goodness-of-fit: the weighted residual sum of squares is chi-square with (m - 2) degrees of freedom when the two-cause model holds, so the p-value map flags pixels where strain and doping alone cannot explain the measured shifts - something no exactly determined two-mode inversion can detect.
• • compare_mode_sets: expected-precision comparison of candidate mode subsets before an experiment.
• • Reduces exactly to the two-mode SeparationModel core at m = 2 (asserted to machine precision).
    Also: scipy declared as an explicit dependency; README Status and Roadmap reconciled with the release history.