feat: multi-fidelity support via Phase-level world/scorer overrides

[jc-macdonald]

Summary

Adds multi-fidelity support: each Phase can optionally override the Study-level world and/or scorer, enabling cheap-surrogate-then-expensive-model workflows (#78).

API change

@dataclass
class Phase:
    name: str
    grid: ...
    filter_fn: ... = None
    n_trials: int = 100
    world: Simulator | None = None   # NEW — None inherits from Study
    scorer: Scorer | None = None     # NEW — None inherits from Study

Backward-compatible — defaults to None, so all existing code works unchanged.

Design rationale

Fidelity is a computational strategy, not a design factor — you don't optimize over it, you use it to allocate compute. A phase-level override keeps fidelity concerns at the architectural level (which phase uses which model) rather than mixing them into the factor space.

Changes

File	Change
src/trade_study/study.py	Phase gains world and scorer fields; Study.run() resolves per-phase overrides
tests/test_study.py	5 new tests: world override, scorer override, both combined, screen-then-validate, default None
examples/bayesian_study.py	Simulator accepts n_samples parameter; new multi-fidelity section (50 draws screen → 2000 draws validate)
docs/guide/bayesian.md	Multi-fidelity section with cross-domain examples (epi, engineering, forecasting)

CI

• 263 tests pass, 99.58% coverage
• Ruff, mypy all clean
• Example script runs end-to-end

Closes #78