[codex] Fix dataset directory loading

[Bortlesboat]

Summary

Adds deterministic directory handling to the shared eval/train JSON record loader. Existing dataset directories now load immediate JSON/JSONL/gzipped record files instead of being passed to open() as if they were files.

Root Cause

iter_json_records() checked os.path.exists(dataset_path) before distinguishing files from directories. When callers passed an existing dataset directory, the directory path was added to file_paths, and the loader raised PermissionError while trying to open it.

Verification

• python -m unittest tests.test_eval_data -v
• python -m py_compile opf/_eval/data.py tests/test_eval_data.py
• No-model loader probe:
  • examples/data/sample_eval_five_examples.jsonl -> 5 records
  • examples/data/*.jsonl -> 10 records
  • examples/data -> 10 records

Note: full-package python -m compileall -q opf on current upstream main is blocked by the existing opf/_train/runner.py syntax error addressed separately in #1.