Add R2DatasetLoader to replace HuggingFace data source

[calm329]

Summary

• Add R2DatasetLoader class that fetches training data from Cloudflare R2 parquet files instead of HuggingFace API
• Add --neuron.data_source config option to switch between huggingface (default) and r2
• Add get_dataset_loader() factory function for seamless loader selection
• Update miner, validator, reward, avg_handler, and eval scripts to use configurable loader
• Maintain same deterministic seeding mechanism (block + UID) for data selection

Test plan

• Set R2 dataset environment variables (R2_DATASET_ACCOUNT_ID, R2_DATASET_BUCKET_NAME, R2_DATASET_ACCESS_KEY_ID, R2_DATASET_SECRET_ACCESS_KEY)
• Run miner with --neuron.data_source r2 and verify data loads from R2
• Run miner with --neuron.data_source huggingface (or default) and verify existing behavior unchanged
• Verify deterministic page selection produces same results for same block + UID

Closes #49

---

Contribution by Gittensor, learn more at https://gittensor.io/