[BUG]: `FileMetaData.c_obj` can be null and segfault on dereference

[Matt711]

FileMetaData has no __cinit__ that default-constructs c_obj, so any instance created via
FileMetaData.__new__(FileMetaData) (bypassing __init__, which raises) has a null c_obj.
Any dereference(self.c_obj) call then segfaults instead of raising a Python exception:

cudf/python/pylibcudf/pylibcudf/io/parquet_metadata.pyx

Lines 473 to 480 in d4ecace

	def __init__(self):
	raise ValueError("FileMetaData cannot be constructed directly")
	@staticmethod
	cdef FileMetaData from_libcudf(unique_ptr[cpp_FileMetaData] metadata):
	cdef FileMetaData result = FileMetaData.__new__(FileMetaData)
	result.c_obj = move(metadata)
	return result

Previously c_obj was an embedded (non-pointer) value and could not be null, so this is a
regression in safety introduced by PR #23558 (which switched FileMetaData to hold a
unique_ptr so it could be shared/reused, e.g. from hybrid_scan.pyx).

Scalar has the same unique_ptr + from_libcudf-only pattern elsewhere in pylibcudf, so this
may be worth fixing generally rather than just for FileMetaData.

Related

• PR [PERF]: Improve the performance of plc.io.parquet.read_parquet with prefetched parquet file metadata. #23558 (introduced the unique_ptr change)
• [PERF]: Improve the performance of plc.io.parquet.read_parquet with prefetched parquet file metadata. #23558 (comment)

[vyasr]

I wouldn't consider this to be a bug. If a class disables __init__ and provides factory methods, and you choose to use __new__, you've chosen a path that at best has sharp edges and at worst is wholly unsupported. If we put work in __new__ we'll be doing extra work for no benefit.