stack: add nns_stack_fit / nns_stack_predict (fit-once, reuse RPM/CV)

Adds a real fit/predict split to NNS.stack so the cross-validated model
selection - the dominant cost - runs once and is reused across predictions,
instead of re-running the full out-of-fold CV on every call.

  * nns_stack gains an opt-in "fitted" bypass: when supplied, the three CV
    fold-loops (dimension count, n.best, ensemble weight) are skipped and the
    stored hyperparameters are used; only the (cheap, test-dependent) final
    production fit + projection runs. The default (fitted=None) leaves behavior
    byte-identical.
  * return_fit=True attaches the selected hyperparameters as the "_fit" state;
    the default public result dict is unchanged (no new key).
  * nns_stack_fit(x, y, ...) runs the selection once and returns a fit object;
    nns_stack_predict(fit, newdata) scores new data reusing it.

Verified: nns_stack_predict(fit, newdata) is byte-identical to a full
nns_stack(x, y, newdata) call across method in {(1,2),(1,),(2,)} times
dim_red_method in {cor, equal, NNS.dep} (the selection is test-independent).
Fit-once + 5 predictions runs ~5.4x faster than 5 full stacks (2000x4,
order='max', folds=5). Full invariant suite (346) passes unchanged.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: claude

src/nns/__init__.py

@@ -65,6 +65,8 @@
     "nns_seas": ("nns.seasonality", "nns_seas"),
     "nns_sd_cluster": ("nns.stochastic_dominance", "nns_sd_cluster"),
     "nns_stack": ("nns.stack", "nns_stack"),
+    "nns_stack_fit": ("nns.stack", "nns_stack_fit"),
+    "nns_stack_predict": ("nns.stack", "nns_stack_predict"),
     "nns_ss": ("nns.stochastic_superiority", "nns_ss"),
     "nns_var": ("nns.var", "nns_var"),
     "prepare_factor_predictors": ("nns.regression", "prepare_factor_predictors"),