config: model-provider selection - #205

Extraction and completion stay behind the existing Extractor/Completer
traits; a new provider module reads GATEWAY_MODEL_PROVIDER and its
settings once at startup, validates them (a named provider with a
missing or malformed setting stops the gateway), builds the client and
owns model routing. An OpenAI-compatible client covers local model
servers (Ollama, vLLM, llama.cpp) with json_schema extraction, refusing
truncated, filtered or non-JSON answers and keeping the upstream status
on errors. ANTHROPIC_API_KEY alone keeps the previous behavior.

Part of #87 (provider selection).