perf: iterate reason codes via to_dict('records') instead of iterrows

[AmirhosseinHonardoust]

add_reason_codes iterated rows with DataFrame.iterrows(), which builds
a pandas Series per row, and reason_codes_for_row then re-wrapped each
row in pd.Series. Both are unnecessary: the logic only uses row.get(),
which dicts support directly.

• Iterate df.to_dict("records") and pass plain dict rows.
• Drop the per-row pd.Series(row) construction; reason_codes_for_row
  now accepts a Series or any mapping unchanged.

No behavior change (O(n) before and after; this is a constant-factor
speedup, not an algorithmic fix).

Verified locally:

• Reason codes identical for Series-input vs dict-input across all 700
  rows of the scored test set; full scored CSV byte-identical (md5
  unchanged).
• ~6x faster on a 20k-row benchmark (0.99s -> 0.17s).
• ruff/black/mypy clean; 34/34 tests pass.