fix: raise TypeError for unsupported partition source type

[anxkhn]

Rationale for this change

The @singledispatch base of _to_partition_representation in
pyiceberg/partitioning.py returned its error instead of raising it:

return TypeError(f"Unsupported partition field type: {type}")

For an unsupported (non-primitive) partition source type, the base handler is the
one that runs, and it handed the TypeError instance back as the partition value
rather than raising it. The exception object would then flow downstream as the
"iceberg-typed" partition value (through partition_record_value ->
PartitionKey.partition -> Record) and only blow up later during serialization,
far from the real cause, or write a garbage partition record.

This is the only return <Exception>(...) in the pyiceberg/ tree. Every sibling
registered handler for this function raises (the TimestampType / DateType /
TimeType / UUIDType / PrimitiveType registrations all raise ValueError), as
does every other singledispatch base handler in the codebase (for example the
to_bytes / from_bytes / to_json / from_json bases in
pyiceberg/conversions.py). So raise is the correct, convention-matching
behavior; the fix is a one-word correction (return -> raise).

In normal use check_compatible blocks non-primitive partition sources upstream, so
this base handler is reached only in edge cases: this is a latent robustness fix with
no behavioral downside on the happy path.

Are these changes tested?

Yes. A focused regression test was added in tests/table/test_partitioning.py
(test_to_partition_representation_unsupported_type) that asserts the base handler
raises TypeError for a non-primitive (StructType) source type. It fails on the
old return (DID NOT RAISE TypeError) and passes after the change to raise.

python -m pytest tests/table/test_partitioning.py \
  -k to_partition_representation_unsupported_type -v
# -> 1 passed

The linters (ruff, ruff-format, mypy and the rest of the pre-commit suite) pass on
both changed files. No dependencies changed, so uv.lock is untouched.

Are there any user-facing changes?

No public API changes. The only observable difference is that an unsupported
partition source type now raises TypeError at its source (as it was always meant
to) instead of silently yielding an exception instance as the partition value.

[ebyhr]

partitioning.py change looks good. I'm not convinced the new test adds much value, though.

The base handler is unreachable in normal use, so it exercises an implementation detail rather than observable behavior.

[ebyhr]

No other same mistake as far as I checked.