Fix HDF5Writer id()-reuse bug corrupting class_index in multi-worker generation

[sclkim1]

Fixes #478.

Problem

HDF5Writer stamps a stable, counter-based _hdf5_key on each Signal object before writing (_assign_hdf5_keys), but not on the signal's metadata .parent chain (the generator/class-level HierarchicalMetadataObject that populate_hdf5_group_with_metadata() recurses into). Those parent objects fell back to str(id(obj)) in _hdf5_key().

Parent metadata objects aren't guaranteed to stay alive for the whole write session - a new one can be built per batch/worker and then garbage-collected once its signals are written. When CPython recycles a freed parent's memory address for a later, unrelated parent object, both resolve to the same key, and populate_hdf5_group_with_metadata()'s dedup guard (if key in group: return False) silently keeps the first parent's data and skips writing the second. Every signal pointing at the second (skipped) parent then reads back the first, unrelated parent's metadata fields on load - most visibly, class_index coming back as a list like [correct_label, wrong_label] instead of a scalar.

This only shows up with num_workers > 1 (real-scale generation via WorkerSeedingDataLoader), and the corruption rate grows with run size: 0% single-process, ~1.4% at 3,000 samples/4 workers, ~67% on a real 1,060,000-sample production run in our case. It is not specific to any signal type or impairment level.

See #478 for the full writeup.

Fix

HDF5Writer._assign_hdf5_keys() now also stamps a stable key on the metadata .parent chain (_assign_hdf5_keys_to_parent_chain), reusing an existing key via hasattr so a genuinely-shared parent object still dedupes to one HDF5 group - only objects without a key yet (i.e. those at risk of an id() collision) get a fresh counter-based one.

Testing

• Added test_writer_no_parent_metadata_id_reuse_corruption (tests/utils/test_writer.py, marked @pytest.mark.full): generates 3,000 samples through the real multi-worker WorkerSeedingDataLoader/DatasetCreator path and asserts no sample's class_index comes back as a list. Passes locally after the fix.
• Manually reran a 50,000-sample/10-worker generation that reliably reproduced the bug pre-fix; 0/50000 corrupted afterward.
• Regenerated a real 1,060,000-sample + 106,000-sample production dataset with the fix; corruption went from ~67% to 0/1,166,000.