feat: Improve compatibility with agno 2.5.0

[Abhijeet Prasad (AbhiPrasad)]

resolves https://linear.app/braintrustdata/issue/BRA-4192/agno-integration-improvements

AI Summary

This PR hardens and modernizes the Agno integration to support the API changes introduced around the Agno 2.4 -> 2.5 transition, while preserving compatibility with older versions.

Why this change was needed

Agno’s run execution model changed significantly:

• older versions relied heavily on private methods like _run/_arun/_run_stream/_arun_stream
• newer versions moved execution behind public dispatcher methods (run/arun) with mixed return contracts (direct value, iterator, async iterator, awaitable returning async iterator)

Our wrapper previously had gaps in these newer async/streaming paths:

• spans could end too early for awaited async-stream dispatchers
• streamed task metrics could be dropped in aggregation for object-like metrics payloads
• duplicated wrapper logic in agent.py/team.py made this behavior harder to maintain safely

What changed

Integration behavior

• Updated public run/arun wrapping logic for both Agent and Team to handle all dispatcher return shapes correctly.
• Fixed async streaming lifecycle bug: spans now remain open until stream consumption completes (no premature span.end()).
• Fixed streamed task metrics aggregation so token metrics are preserved even when metrics are object-like.

Refactor / maintainability

• Extracted shared public dispatch wrapping logic into:
  • py/src/braintrust/wrappers/agno/run_helpers.py
• Reduced duplication between Agent and Team wrappers.

Tests

Expanded py/src/braintrust/wrappers/test_agno.py with coverage for:

• public dispatcher streaming (sync + async)
• awaited async iterator dispatch path
• non-stream async dispatch path
• sync/async stream error paths
• early-break stream consumption paths
• parent span nesting
• private-method precedence over public dispatch methods
• strict span lifecycle regression checks

Also added short explanatory docstrings for complex tests.

Version matrix

• Updated Agno nox matrix to:
  • LATEST, 2.4.0, 2.1.0
• Added comment explaining this covers newest API plus the key 2.4→2.5 transition boundary.

User-facing changes

• Positive behavior change for users on newer Agno versions:
  • more reliable tracing in async streaming paths
  • correct span timing/lifecycle for dispatcher-based arun(..., stream=True)
  • improved streamed token metrics capture in task spans
• No breaking API changes to Braintrust wrapper usage (setup_agno(...) remains the same).