fix: handle nullable evaluator diagnostics

[Letian Liu (0x90000)]

Summary

Fixes #1458.

CoSTEERSingleFeedback.return_checking is nullable, but the data-science workflow and pipeline evaluators append validation diagnostics with +=. When model feedback leaves this field as None, a failed deterministic check raises TypeError instead of returning structured negative feedback.

This change adds a shared null-safe append helper and uses it for all workflow/pipeline validation diagnostics. Existing text is preserved and None is initialized to the same diagnostic format as before.

Tests

• Added regression tests for both None and existing return_checking values.
• Python compilation passes for all changed files.
• Ruff import sorting and formatting pass for the new test; the repository's existing evaluator files contain unrelated baseline lint violations.
• Full pytest collection is blocked in this Windows environment because the optional dill dependency is not installed; a direct helper behavior check passes.

---

📚 Documentation preview 📚: https://RDAgent--1480.org.readthedocs.build/en/1480/

[Letian Liu (0x90000)]

@microsoft-github-policy-service agree