Harden the v0.1.0 release

[tutao0123]

Summary

• add Python 3.10/3.12 CI, documentation consistency tests, shell checks, and a real offline CPU end-to-end smoke run for AR and MDLM
• separate runnable configs from the non-runnable 350M FineWeb architecture target and remove the stale TinyStories 50M config
• add release metadata and contributor documentation (CHANGELOG.md, CITATION.cff, CONTRIBUTING.md, project URLs)
• update the English and Chinese READMEs with status badges, CPU smoke instructions, and config policy

Impact

This makes the current diffusion-language-model trainer and GPT-style AR trainer suitable for a first tagged public release without implying that Block Diffusion, pretrained checkpoints, or the FineWeb data recipe already exist.

Local validation

• ruff check .
• pytest (22 passed)
• bash -n scripts/*.sh
• bash scripts/run_cpu_smoke.sh (AR and MDLM both trained, checkpointed, reloaded, and sampled on CPU)
• git diff --check

No datasets, checkpoints, generated outputs, or API keys are included.