Improve run_dataloader.py usability and performance feedback(addresses #869)

[ezylopx5]

Summary

Improves run_dataloader.py usability for users who need to extract partial training data or debug data loading performance (addresses #869).

Changes

• Add --max_batches flag to limit processing (useful for testing or partial data extraction)
• Add --local_data_root flag to auto-substitute remote paths with local cached data
• Add real-time throughput statistics (batches/s, MB/s) to progress bar
• Warn users when loading from remote paths without local cache
• Fix: Save any remaining batches that don't fill a complete file (was silently dropping data)
• Log final statistics at completion

Motivation

Users running run_dataloader.py with remote data (S3/R2) experience very slow performance due to per-chunk network requests (#869 reported only 44 batches after running overnight).

This PR makes the script more user-friendly by:

1. Allowing partial runs with --max_batches for testing
2. Making it easier to use local data with --local_data_root
3. Providing clear feedback about processing speed

Example usage

# Process only first 100 batches
python scripts/run_dataloader.py -o ./output --max_batches 100 config.yaml

# Use locally cached data instead of S3
python scripts/run_dataloader.py -o ./output --local_data_root /mnt/data config.yaml