LMFit tolerance routed by raw method argument instead of configured enum method

[rozyczko]

Finding F-H5 in DEEP_ANALYSIS.md (parent #261). Verified empirically.

minimizer_lmfit.py:162, 246-249 — _get_fit_kws(method, ...) receives the raw method argument, which is None on the normal enum-configured path, so tolerance is always mapped to ftol:

• LMFit_cobyla + fitter.tolerance -> TypeError: minimize() got an unexpected keyword argument 'ftol', wrapped in FitError.
• For powell / differential_evolution the intended 'tol' branch is unreachable; the user's tolerance is silently dropped or misapplied.

Fix: key the mapping off self._method (the resolved method), not the raw argument.