feat: Add warning for empty or sentinel LLM responses

[RMtz88]

Summary

Detect and warn when model returns empty content or sentinel tokens like <|endoftext|> without tool calls or refusals.

Fixes #376

Changes

• Add _warn_empty_response() helper method (10 lines)
• Call helper in get_response() and stream_response()
• Add 3 unit tests

Behavior

When the model returns an empty or sentinel response, users will see:

"Model completed without output (sentinel). May indicate model-agent compatibility issue."

Notes

• Non-breaking: warning only, no change to control flow or return types
• No new dependencies
• Follows existing logging conventions