Improve NG init-robustness on under-determined components (best-of-N restarts)

[neuromechanist]

Follow-up to #145 (resolved as init-basin sensitivity, not a dynamics bug).

The same-init test showed AMICATorchNG and Fortran agree from an identical init (0.9974 mean, 0/70 collapsed on the full 70ch/k=152 data), but torch is more sensitive to certain random inits than Fortran: a specific seed (42) lands ~7-10 of the weakest, under-determined components in a different (equal-or-higher-LL) basin than Fortran's, at the full 2000-iter do_newton budget. Fortran is init-robust (0.9999 self-consistency from any random init).

This is not a parity bug (the M-step / Newton are algebraically identical to amica15.f90; identical init -> matching answer). But torch's single-fit robustness could be improved so a run is less likely to land in a seed-42-style basin:

• Best-of-N restarts: run K seeds, keep the highest-LL (or most-reference-consistent) fit. Extends the Multi-model NG log-likelihood is ~0.02 lower and more variable than Fortran #51 keep_best idea across seeds.
• A more robust default initialization.

Not blocking: the do_newton=0 single-model conformity (0.998) is the reproducible parity claim. Full analysis in .context/issue-145/same_init_result.md.