Normalize corpus to pure ASCII (transliterate non-ASCII, drop emoji/control bytes)

Add all.txt: single-file concatenation of all corpus files for LLM training