[codex] Fix data parser cache leaks with attribution

[ObservedObserver]

Summary

• Preserve the original Fix memory leak by replacing @lru_cache with instance-level caching in data parsers #724 commits from Lilo so the initial fix authorship is retained.
• Replace method-level @lru_cache() on data parser cached properties with per-instance cached_property in the final implementation.
• Add regression coverage proving pandas parser cached properties do not retain parser/DataFrame instances and that parsers remain picklable after cache loading.

Root Cause

functools.lru_cache stores calls in the decorated function wrapper. For instance methods and properties, that cache key includes self, so parser instances are retained by the class-level cache after callers drop their references. Those parser instances keep their DataFrames alive, causing memory growth in long-running processes that parse many datasets.

Notes

This PR carries forward the original work from #724, then adds a compatibility adjustment: the final implementation uses cached_property instead of storing a threading.Lock on parser instances, so parser pickling keeps working.

Supersedes #735.
Fixes #723.

Validation

• ./venv/bin/python -m pytest tests/test_data_parsers.py -q
• ./venv/bin/python -m pytest -q
• ./venv/bin/python -m ruff check .