fix(io): write NaN values into identity-partitioned float columns

[breken-ai]

Rationale for this change

Appending a NaN to a table with an identity partition on a double or float column fails:

schema = Schema(NestedField(1, "id", IntegerType()), NestedField(2, "value", DoubleType()))
spec = PartitionSpec(PartitionField(source_id=2, field_id=1000, transform=IdentityTransform(), name="value"))
tbl = catalog.create_table("default.t", schema=schema, partition_spec=spec)
tbl.append(pa.table({"id": [1, 2, 3], "value": [1.0, float("nan"), None]}))
# ZeroDivisionError: division by zero  (bin_pack_arrow_table: tbl.nbytes / tbl.num_rows)

_determine_partitions groups the rows by partition value, which yields one NaN group, and then selects each group's rows with pc.field(name) == value. NaN never compares equal to itself, so the NaN partition selects zero rows and the empty table reaches bin_pack_arrow_table. Nulls are already handled with is_null(). This PR selects the NaN partition with pc.is_nan() in the same way.

After the fix the NaN rows are written to their own partition, and value is nan returns them.

Are these changes tested?

Yes.

• tests/io/test_pyarrow.py: test_determine_partitions_identity_nan checks that the rows for 1.0, NaN and None land in their partitions. On main it returns {'nan': []}.
• tests/catalog/test_catalog_behaviors.py: test_append_nan_to_identity_partitioned_table appends to a real table (memory and SQL catalogs), then checks that a full scan returns all 4 rows and value is nan returns the two NaN rows. On main the append raises ZeroDivisionError.

All 4 new tests fail on main and pass with the fix. tests/io/test_pyarrow.py and tests/catalog/test_catalog_behaviors.py pass. prek run --files (ruff, ruff-format, mypy, pydocstyle, codespell) passes. I also checked float columns by hand.

Are there any user-facing changes?

Yes. append()/overwrite() now write NaN values into identity-partitioned floating-point columns instead of failing.

AI disclosure: this bug was found, fixed and tested by an AI coding agent (Claude) running under the breken-ai account; the red/green runs above are its local results.

[Fokko]

This is indeed a bug, thanks @breken-ai for fixing this 👍