Replace heuristic optuna_style search with real trial-based optimization

[sybapp]

Follow-up from review

The current optuna_style mode evaluates an initial prefix, then ranks remaining bounded candidates by similarity to the best survivor plus seeded randomness. That is deterministic and bounded, but it is not a real trial/sampler optimization loop.

What to build

Replace or extend optuna_style with a real trial-based optimizer over Strategy Spec parameters and bounded template choices.

Acceptance criteria

• The optimizer exposes trial/objective semantics for Strategy Spec parameters or bounded template choices.
• Trial results are scored only through schema-valid Strategy Specs and Nautilus walk-forward validation.
• The optimizer records sampler/configuration, trial parameters, objective values, rejected candidates, and winning artifacts in the Run Registry.
• The implementation is deterministic under a configured seed.
• Tests demonstrate a small fixture optimization that improves or selects the best candidate reproducibly.
• LLM-proposed ideas still cannot bypass Strategy Spec validation or Nautilus scoring.

Context

Review finding: current optuna_style is a heuristic ordering mode, not comparable to Optuna-style optimization.

Related issue requirement: "Optuna-style optimization can search Strategy Spec parameters or bounded template choices."