Polish the public GitHub and Hugging Face documentation for CetinLM.
- Update the project overview with the current architecture and corpus
- Add model family and hardware philosophy
- Improve corpus, training, evaluation, and engineering documentation
- Add third-party data provenance and licensing references
- Refresh GitHub and Hugging Face README presentation
- Preserve the existing training system and project structure
fix(data): stabilize Phase-II v5.2 corpus acquisition and validation
* pin all dataset revisions to immutable commit SHAs
* replace fragile lazy HTTP streaming with adaptive Parquet acquisition
* use synchronous row-group reads for normal sources
* use pinned complete-file local batching for coarse Parquet layouts
* persist quality-approved source caches with resume support
* enforce a network-off barrier during final corpus construction
* preserve EN 38% / TR 30% / multilingual 32% training mix
* enforce whole-document validation with hard source-diversity reservation
* fix Turkish validation starvation by reserving capacity for missing sources
* retain boundary, content-quality, domain-diversity, and near-dedup gates
* verify zero transport errors and zero HF retries in the 1M pilot
