`DataFilesList.filter()` leaves `origin_metadata` misaligned with the filtered files

Description

DataFilesList keeps origin_metadata, a list parallel to the file list where each entry describes the origin of the file at the same index (e.g. the data files pattern it was resolved from). After filter() removes some files, the returned DataFilesList keeps the unfiltered origin_metadata, so metadata no longer lines up with the files it describes: len(origin_metadata) is larger than the number of files, and origin_metadata[i] refers to a different file.

Reproduction

from datasets.data_files import DataFilesList

# two resolved files, each with origin metadata
data_files = DataFilesList.from_patterns(["*"], "/path/to/dir/with/data.txt/and/data.csv")

filtered = data_files.filter(extensions=[".txt"])
assert len(filtered.origin_metadata) == len(filtered)  # fails: 2 != 1

Any consumer iterating zip(filtered, filtered.origin_metadata) or indexing filtered.origin_metadata[i] associates the wrong metadata with a file, and indexing past len(filtered) reads stale metadata.

Expected behavior

filter() should apply the same filtering to origin_metadata so the one-to-one alignment between files and their metadata is preserved.

Environment

• datasets version: 5.0.2.dev0 (main, d336dcb)
• Python: 3.11