[Data] Add include_row_hash to read_parquet

[wingkitlee0]

Description

This PR adds an include_row_hash option to read_parquet.

Row hash is unique for each row across the whole dataset. This can be used for checkpointing (for Ray Data and/or Ray Train pipeline)

Related issues

Closes #61410

Additional information

How it works:

1. Path seed: For each Parquet file, MD5-hash its file path and take the first 8 bytes as a uint64 seed. This means
   identical data in different files always produces different hashes.
2. Row keys: Add the row's position within the file (0-based offset tracked across batches) to the path seed: key =
   path_seed + row_index.
3. Mix: Apply the splitmix64 finalizer (a bijective 64-bit integer mixing function) to scatter nearby keys across
   the full int64 range:

  keys ^= keys >> 30
  keys *= 0xBF58476D1CE4E5B9
  keys ^= keys >> 27
  keys *= 0x94D049BB133111EB
  keys ^= keys >> 31

All operations are vectorized numpy — no Python loops.
Properties:
• Reproducible: Same file path + same row position always yields the same hash.
• Unique: Different files get different seeds (via MD5 of path); different rows within a file get different
offsets. The splitmix64 finalizer is bijective, so distinct inputs never collide.
• Fast: One MD5 call per file, then pure numpy vectorized arithmetic per batch.

[gemini-code-assist]

Code Review

This pull request introduces a useful include_row_hash option to read_parquet, which is valuable for checkpointing and data versioning. The implementation is generally solid and consistent with existing features like include_paths. However, I've identified a critical bug that can cause a crash when include_row_hash=True is used on a file that already contains a row_hash column, particularly when no specific columns are selected for reading. I've provided details and a suggested fix for this issue. Additionally, I've included a few medium-severity suggestions to improve user experience by adding a warning for column name conflicts, updating the documentation to clarify this behavior, and enhancing test coverage for this edge case.