feat: introduce warm start

[Mateusz-Switala]

Description

Improves the warm-start strategies for the GAM-based HPO optimizer (GAMOptimizer) so greedy and balanced
starts provide meaningful discrete-parameter coverage before GAM-guided search begins.

Motivation

The previous warm-start selection could under-cover discrete parameter values, causing early trials to
cluster and leaving the fitted GAM with sparse categorical observations.

Changes

• Coverage-aware warm starts: Greedy and balanced strategies prioritize combinations that cover
  previously unseen discrete values.
• Balanced strategy: Uses round-robin selection across fields_to_balance value tuples while also
  prioritizing coverage of non-balanced discrete fields.
• Automatic warm-start sizing: For greedy and balanced strategies, the effective warm-start target
  expands to the minimum required for coverage rather than failing or silently under-covering when
  n_random_nodes is too small.
• Evaluation allocation: max_evals is optional; when specified, output allocation accounts for the
  effective warm-start target and reserves remaining capacity for GAM-guided evaluations.
• Model-value handling: Normalizes model objects, dictionaries, and strings for coverage tracking;
  serializes model-valued DataFrame columns to model_id before GAM fitting.
• Balanced-selection fix: Corrected the ordering/fallback behavior so round-robin balancing is
  preserved.
• Tests: Expanded test_gam_opt.py for coverage, round-robin behavior, model serialization, auto-
  adjusted warm starts, and post-warm-start GAM iterations.

Testing

• tests/unit/ai4rag/core/hpo/test_gam_opt.py

Checklist

• Tests added/updated
• Documentation updated
• Code follows style guide
• All checks passing