feat: detect context window limit errors for auto-compact

[jhult]

Add ContextLimitError class to detect when upstream API returns context window limit errors. The error response includes a suggest_compact flag that clients can use to trigger their own conversation compaction.

When a context window limit error is received from the upstream API, the proxy now automatically compacts and retries for non-streaming requests:

1. Generates a structured summary with 5 sections: Task Overview, Current State, Important Discoveries, Next Steps, Context to Preserve
2. Places summary as a single assistant message (replacing entire message history)
3. Increases max_tokens to 2000 for detailed summaries
4. Requests <summary></summary> tag wrapping
5. Retries the original request with the compacted conversation

This mimics client-side /compact command behavior, allowing conversations to continue seamlessly when hitting token limits.

Auto-compact retry only applies to non-streaming requests since streaming responses cannot be retried mid-stream.