Fix is_dollar_amount() to accept strings cast from single precision floats. Ensure no mixed dtypes before to_parquet()

[radkoff]

Two fixes here

1. This issue
2. Some dollar amounts are cast to strings from single-precision floats (e.g. 10.5 instead of 10.51), and the existing is_dollar_amount does not recognize it. This means a perfect match is not detected in dollar_match(), which was causing my results to be much lower than expected. I saw it happening in your baseline results too (❌ gross_amount: guessed "$1,510.00" with score 0.903, was actually "$1,510.0" ), so if you guys pull this and rerun, you should see a higher score for gross_amount