fix: avoid summing cumulative streaming usage on OpenAILike models - #11549

Summary

Fixes #11538

OpenAI-compatible servers like vLLM (with force-include-usage or continuous usage stats) emit running cumulative token counts on intermediate SSE chunks and a final usage summary chunk with an empty choices list. Because OpenAILike inherited collect_metrics_on_completion = False, _populate_stream_data added every chunk's cumulative usage together, and enabling collect_metrics_on_completion still dropped the final choices: [] usage frame because _should_collect_metrics returned False when response.choices was empty.

This sets collect_metrics_on_completion: bool = True by default on OpenAILike, updates OpenAIChat._should_collect_metrics to accept the final empty-choices usage chunk, and replaces (while preserving timer, duration, and TTFT) rather than adds stream_data.response_metrics when collect_metrics_on_completion is enabled so streams that attach usage to both the finish_reason chunk and the trailing choices: [] chunk do not double-count.

Type of change

• Bug fix

Checklist

• Code complies with style guidelines
• Ran format/validation scripts (./scripts/format.sh and ./scripts/validate.sh)
• Self-review completed
• Documentation updated (comments, docstrings)
• Tested in clean environment
• Tests added/updated (if applicable)

Duplicate and AI-Generated PR Check

• I have searched existing open pull requests and confirmed that no other PR already addresses this issue

Additional Notes

Verified with pytest libs/agno/tests/unit/models/openai/test_openai_metrics.py libs/agno/tests/unit/models/perplexity/test_perplexity_metrics.py (23 passed).