Fix/563 eval model dict csv

[prawnsgupta]

Fixes #563.

Problem. pm_evaluate ... --out_fname results.csv crashes with AttributeError: 'dict' object has no attribute 'to_csv' for detection models. model.eval() returns a dict {"metrics_df", "metrics_factory"} for detection (torch_detection.py), while segmentation/other tasks return a DataFrame; cli/eval_model.py called .to_csv() directly on the result. (The issue referenced cli/evaluate.py:197, but the CLI has since been refactored into eval_model.py / eval_preds.py; the bug is now at eval_model.py:209.)

Fix. Extract the metrics DataFrame with results.get("metrics_df") if isinstance(results, dict) else results — the same pattern already used in cli/batch.py — before writing the CSV, and preserve the function's existing return results contract. Also added os.makedirs(os.path.dirname(os.path.abspath(out_fname)), exist_ok=True) to match cli/eval_preds.py, so --out_fname with a non-existent parent directory works.

Scope. cli/eval_preds.py uses a different path (dataset.eval_preds(), which returns a DataFrame) and already creates the directory, so it needs no change; cli/batch.py already handled the dict case.