SDK Core 0.3.6

Fixed

• Export trace spans with the worker resource (service.name=agnt5-worker, service.version, agnt5.*). The span filter did not forward the provider resource to the OTLP exporter, so Python and TypeScript traces had an empty resource and never appeared in trace listing. The filter now also forwards force_flush and shutdown_with_timeout.
• Treat the gpt-6 family as OpenAI reasoning models, like gpt-5 and the o-series: no temperature or top_p in the request and max_completion_tokens instead of max_tokens. gpt-6 rejects both with a 400, so every Python call to gpt-6-luna failed.