[DG47]

Closes #4130

Rationale for this change

Parquet statistics do not track NaN values, so data files written by PyIceberg never carry nan_value_counts. The strict metrics evaluator reads a missing NaN count as "no NaN values" (as Java does, since Java's Parquet writers always record these counts). That means Table.delete("x < 5.0") drops a file containing [1.0, NaN] as a whole, so the NaN row is lost even though it does not match the filter.

This change counts NaN values of float and double columns in ParquetFormatWriter.write, including nested struct fields, list elements and map values. It skips columns whose metrics mode is none, matching how the Java writers populate nan_value_counts. The counts are passed to data_file_statistics_from_parquet_metadata through a new optional argument. Files without float or double columns do no extra work.

Files that are already written keep their missing counts. Files registered through add_files are unchanged, since that path only has the Parquet footer.

Are these changes tested?

Yes.

• tests/io/test_format_writers.py: NaN counts for top-level float and double columns, a struct field, list elements and map values across multiple write() calls, zero counts for columns without NaN, and no count for a column with metrics mode none.
• tests/catalog/test_catalog_behaviors.py: delete("value < 5.0") on a file with [1.0, NaN] keeps the NaN row (fails on main for the memory and sql catalogs).
• tests/integration/test_writes/test_writes.py::test_merge_manifests: updated the expected nan_value_counts from [] to [(6, 0), (7, 0)] for the float and double columns.

make test and make lint pass locally.

Are there any user-facing changes?

Data files written by PyIceberg now include nan_value_counts for float and double columns. This fixes the delete above and lets scans prune on NaN counts.