fix(quickadapter): isolate holdout from continual state (#156)

fix(quickadapter): isolate holdout from continual state (#156)

Cold-start HPO trials and the pre-refit selection model whenever `test_size`
enables two-stage selection, so `holdout_rmse` is an independent diagnostic;
only the final XGBoost/LightGBM refit continues from the FreqAI-supplied
previously deployed model. Other regressors ignore any prior model state.

Give the changed HPO objective a stable semantic identity
(`candidate-cold-start-v1`) stored as the Optuna study `user_attr`
`objective_identity` and in the persisted best-params envelope, so legacy
warm-state studies/params are ineligible (reset/rejected once on upgrade).
Unify the hp and label study-lifecycle versioning behind one
`_OptunaStudyMarker` dispatch (stable study names; hp always resets on
mismatch, no new tunable; label keeps `reset_label_study_on_schema_mismatch`).

Make the canonical `continual_learning=false` explicit; document the
`continual_learning` and `test_size` rows (`0` = single-stage, train_test_split
only) and harmonize strategy comment code refs to RST double backticks.

Fixes #132