fix(runners): unify provider resolution plan for dispatch, hashing, and provenance

[BerryUIKI]

Closes #110.

Overview

Unifies provider resolution and execution planning in CreativeRunner:

• Introduces resolve_execution_plan(req) -> CreativeExecutionPlan as the single source of truth for execution planning, capability validation, dispatch routing, cache key calculation, and provenance recording.
• Eliminates divergent heuristics and silent provider switching between resolver, cache key generation, and execution dispatch (e.g. engine_id='fal_ai' with model='sdxl-turbo' will strictly target Fal.ai without falling back or switching to SiliconFlow).
• Rejects unsupported provider and action combinations immediately with clear actionable errors instead of fallback deviations.
• Attaches resolved provider_id to both compute_creative_cache_hash and GenerationProvenance.
• Adds unit test suite in backend/tests/test_provider_resolution.py verifying the exact issue reproduction scenarios and plan enforcement.