docs: add guide for running the service locally as an API endpoint

[smirnovlad]

Adds docs/service/running_locally.md — a general guide for building the ThinkBooster service locally and calling it as an OpenAI-compatible endpoint, with the scaling strategy and scorer in the URL.

It leads with the no-GPU path (self_consistency against a remote model), then covers beam search / best-of-N with the uncertainty scorers, and folds the PRM scorer and local generation into one GPU section (env vars, Docker GPU override, the tts_api_key remote-generation routing). Cross-links api_guide.md for the full parameter and metadata reference.

This replaces the earlier PRM-specific draft on the branch — the guide is now general and not PRM-named. The launch article (#258) links here for the "run it as a service" path. All referenced files (docker-compose.yml, service_app/Dockerfile, .env.example, the [service] extra) are verified present on main.