fix: security and reliability improvements

[hidirektor]

Summary

• C-1: Replace bare except infinite loop in interaction_dataset.py with bounded retry (max 30) and specific exception types
• H-1: Run full test suite in CI (pytest tests/) instead of single file
• H-2: Fix gsutil subprocess command injection risk in waymo_utils.py — strip newline, validate path, use list args
• H-3: Replace pickle with JSON for agent index cache (df_cache.py); legacy .pkl files still load for backwards compat
• M-1: Replace 29 print() calls with logging across 13 files