Improve embedding data loading robustness and dataset structure

[clemsgrs]

Summary

• add worker-local cache support to legacy TileDataset
• add speed.use_parquet toggle (legacy .npy dataset vs parquet catalog dataset)
• switch temp shard format from HDF5 to .npy memmaps to avoid HDF5 lock/permission failures
• merge rank shards with indexed fill (order-independent, bounded memory)
• improve RegionUnfolding repr
• move TileCatalogDataset class to data/dataset.py; keep tile_catalog.py focused on catalog utilities
• clean up tmp_feature_shards at end of run
• force out-of-order DataLoader delivery when supported (in_order=False)
• fix process_list error/traceback dtype handling and stale error clearing on success

Tests

• python -m pytest -q -o addopts='' tests/test_regression_bugfixes.py
• python -m pytest -q -o addopts='' tests/test_tile_catalog.py