provider: fix Mistral 422 (#261) and make SSE idle timeout configurable (#196)

#261: Mistral strictly enforces the OpenAI schema and 422s on the
non-standard reasoning_content message field and top-level thinking
request field. Add strict_openai_schema_endpoint() (matches the mistral
profile id or *.mistral.ai api_base) and suppress both reasoning_content
injection and the thinking field for those endpoints.

#196: Replace the hardcoded 180s SSE_CHUNK_TIMEOUT with a configurable
[provider] stream_idle_timeout_secs (default 180), overridable via
JCODE_STREAM_IDLE_TIMEOUT_SECS, so slow reasoning models that think
silently for minutes don't trip a premature timeout.