[codex] Add original ChatBot-10B architecture

[DoRmAmMu1997]

Summary

• Adds an original untrained ChatBot-10B dense decoder config, not a fine-tune of external model weights.
• Replaces the simple Transformer internals with RMSNorm, RoPE, grouped-query attention, SwiGLU, tied embeddings, and richer generation controls.
• Adds BPE tokenizer tooling, dataset manifests/loaders, tiny configs, parameter estimation, tests, CI, and updated training docs.

Test Plan

• python -m compileall chatbot.py chat_llm.py train_llm.py train_tokenizer.py train_10b.py src scripts
• pytest -q (11 passed locally)
• python scripts\estimate_params.py --config configs\chatbot-10b.yaml (reports 9.999B params)
• git diff --check

Notes

• Full 10B weights are intentionally not committed.
• Full 10B training requires external multi-GPU/cloud infrastructure; CI validates the tiny CPU path.

Co-authored-by: Codex codex@openai.com