eea19d2: Add support for the context_usage server event via a new onContextUsage callback. The event is emitted after each completed agent turn and reports { event_id, model, context_tokens, context_limit_tokens }, where context_tokens is the prompt size of the turn's last LLM generation and context_limit_tokens is that model's maximum context window - useful for surfacing how close a long conversation is to the context limit.
