Fix HLLM embedding/vocab alignment; add generative data download workflow

[TyndaleLym]

Pull Request / 拉取请求

What does this PR do? / 这个PR做了什么？

Fix HLLM item-embedding/vocab alignment so embeddings are indexed by vocab token id (row 0 = PAD). HLLMModel now validates item_embeddings.shape against (vocab_size, d_model), L2-normalizes item embeddings once at load and normalizes the user representation at scoring, so logits = cos-sim(x, emb) / temperature with default temperature=0.07 to match the official ByteDance HLLM logit scale. MovieLens-1M and Amazon Books HLLM preprocessing scripts now read vocab.pkl and write embeddings keyed by token id; the example run scripts compute vocab_size from item_to_idx and use temperature=1.0 in the NCE loss to avoid double scaling. Also fix two Amazon Books data bugs: the interaction loader now coerces item_id/user_id to str so the vocab built from numeric ByteDance ids lines up with the str-keyed metadata used by the embedding-generation lookup (previously this silently produced an all-zero embeddings tensor under the default --data_source bytedance), and the ByteDance interaction filename is split from ratings_Books.csv to amazon_books_interactions.csv with a schema-vs-source mismatch error so a stale raw download cannot silently be re-used. Add a unified dataset download workflow with --no_download / --data_source / --overwrite CLI flags on the MovieLens-1M and Amazon Books preprocessing scripts, supporting Stanford SNAP raw data and ByteDance HLLM preprocessed data, with progress bars, automatic format detection and basic schema validation. Add smoke tests covering HLLM vocab/embedding validation, bounded cosine logits, shape error messages, numeric item_id coercion, str-keyed vocab/metadata lookup and schema-mismatch handling, and update the HLLM reproduction docs (EN/ZH) and data README files accordingly.

Type of Change / 变更类型

• 🐛 Bug fix / Bug修复
• ✨ New model/feature / 新模型/功能
• 📝 Documentation / 文档
• 🔧 Maintenance / 维护

How to Test / 如何测试

# Correctness smoke tests
python -m pytest tests/test_hllm_embedding_alignment.py tests/test_amazon_books_preprocess.py -v

# MovieLens-1M data preparation
python examples/generative/data/ml-1m/preprocess_ml_hstu.py
python examples/generative/data/ml-1m/preprocess_hllm_data.py --model_type tinyllama --device cuda

# Amazon Books data preparation (ByteDance preprocessed source, default)
python examples/generative/data/amazon-books/preprocess_amazon_books.py --data_source bytedance
python examples/generative/data/amazon-books/preprocess_amazon_books_hllm.py --data_source bytedance --model_type tinyllama --device cuda

Checklist / 检查清单

• Code follows project style (ran python config/format_code.py) / 代码遵循项目风格（运行了格式化脚本）
• Added tests for new functionality / 为新功能添加了测试
• Updated documentation if needed / 如需要已更新文档
• All tests pass locally / 所有测试在本地通过