Add a robust, faster fitting method to lagFitVelocity()

[AstroEloy]

Summary

lagFitVelocity() (wmpl/Utils/AlphaBeta.py) previously had a single, undocumented fitting path: scipy.optimize.basinhopping(niter=200) + Nelder-Mead on the exact L1 lag residual, seeded from hardcoded magic-number initial guesses, with dead commented-out code and no input validation. This PR:

• Adds an opt-in method='robust' — scipy.optimize.least_squares(loss='soft_l1') with a small multi-start over t0 (the one parameter that makes the exponential→linear model's cost landscape multimodal), instead of 200 global basinhopping restarts.
• Measured ~15-18x faster, and — measured, not assumed — as good or better fit quality: across 3 distinct parameter scenarios × 3 noise realizations (deterministic, fixed seeds), median RMSE-to-truth was 0.61 m/s for robust vs 1.89 m/s for basinhopping; mean 0.96 vs 1.93; robust won outright in 6/9 trials. basinhopping's fixed initial guess + finite iteration budget occasionally leaves it under-converged when the true a1 is far from that guess — least_squares doesn't share that failure mode.
• brentq was considered and confirmed not applicable: unlike alphaBetaVelocityNormed() (which inverts height↔velocity), expLinearLag()/expLinearVelocity() are already explicit closed-form functions of t — there's nothing to invert here.
• Fully backwards compatible: default stays method='basinhopping', byte-for-byte unchanged behavior; the old 4-positional-argument call signature still works.
• Added input safeguards: length mismatches, unknown method, too few points, non-finite/non-positive v0, and degenerate (zero-span) time_data all raise ValueError; non-finite (time, lag, vel) points are dropped with a warning instead of silently propagating NaN.
• Wrote a full docstring in the module's established style (Arguments/Keyword arguments/Return/caveats), removed dead commented-out code (an unused curve_fit fallback, a stale debug-plot block).
• Wired a new -r/--lagrobust CLI flag into the if __name__ == "__main__": block, mirroring the existing -e/--errors pattern — when set, the lag-smoothing step uses method='robust' instead of the default.

Testing

Added 5 new tests to wmpl/Utils/Tests/test_AlphaBeta.py:

• testLagFitVelocityBackwardsCompatibleCall — old positional-only call still works, defaults to basinhopping.
• testLagFitVelocityRobustRecoversParams — robust recovers true params and is faster than basinhopping.
• testLagFitVelocityRobustFitQuality — deterministic multi-scenario Monte Carlo proving robust matches or beats basinhopping's fit quality (not just speed).
• testLagFitVelocityInputValidation — all new safeguards raise ValueError.
• testLagFitVelocityDropsNonFinitePoints — non-finite points are dropped, not propagated.

Full suite: 49/49 passing (python -m wmpl.Utils.Tests.test_AlphaBeta), ~2 min total.

Files changed

• wmpl/Utils/AlphaBeta.py
• wmpl/Utils/Tests/test_AlphaBeta.py