feat(learning): learner hyperparameter config and persisted metrics

[salim4n]

Make learner configuration and metrics explicit enough for Studio controls and reproducibility.

Acceptance criteria:

• learner configs are serializable and persisted in run configs;
• metrics expose stable names for Studio panels and CI gates;
• invalid hyperparameters fail before a run starts;
• docs explain default configs for baseline and neural learners.