Speed up UniversalDataset batching for tabular data

[tippered1-debug]

Summary

This PR adds a batched __getitems__ fast path for UniversalDataset when no tokenizer is used.

For tabular-only data, PyTorch DataLoader can request a whole batch of indices via __getitems__. The previous path fetched every row separately through __getitem__, then rebuilt the batch in collate_dict. The new path slices NumPy arrays once per batch and lets collate_dict handle an already batched dictionary.

The tokenizer/text path keeps the old row-wise behavior.

Changes

• Add UniversalDataset.__getitems__ for tabular-only batched fetching.
• Extend collate_dict to accept both list[dict] and already batched dict.
• Make collation dtype-aware using _dtypes_mapping.
• Add focused tests for batched fetching, dtype preservation, and tokenizer fallback.

Benchmark

Synthetic benchmark, 200k rows:

• tabular_balanced: 15.239x faster batch fetching
• tabular_many_cat: 11.083x faster batch fetching
• tabular_large_batch: 6.944x faster batch fetching

This benchmark measures the UniversalDataset / DataLoader batch fetching path, not full model training time.

Tests

• pytest tests/unit/test_text/test_universal_dataset.py -q