[Data] Generate sortable, collision-resistant Dataset IDs

[yuhuan130]

Description

This PR replaces counter-based Dataset IDs with 22-character, Base62-encoded ULIDs.

Each ULID contains a timestamp and random data, making it sortable by creation time and practically collision-free.

The full ID format remains:

{dataset_name}_{dataset_ulid}_{run_index}

This PR also:

• Generates a new Dataset ULID during deserialization.
• Keeps the run index for repeated executions and training epochs.
• Adds RAY_DATA_USE_LEGACY_DATASET_IDS as a temporary compatibility option.

Tests

Focused Dataset ID tests: 5 passed

Also updates test_dataset_id_train_ingest test in python/ray/data/tests/test_stats.py to verify that the same generated Dataset ID is used across epochs while the run index increases.

---

Closes #65073

[gemini-code-assist]

Code Review

This pull request introduces time-sortable, practically unique Dataset ULIDs encoded in Base62, replacing the legacy counter-based Dataset IDs by default, while providing a configuration option to fall back to the legacy behavior. The review feedback focuses on improving robustness and compatibility, specifically by replacing parenthesized context managers in tests to maintain Python 3.8 compatibility, optimizing the Base62 encoding function to avoid string concatenation in a loop, and using getattr when accessing use_legacy_dataset_ids to prevent potential AttributeErrors during initialization and deserialization.