More (useless) code point normalization

[Artoria2e5]

• Big5-HKSCS EUDA→PUA: https://github.com/stanfordnlp/CoreNLP/wiki/Chinese-Private-Use-Area-code-points (crappy script indeed)
• Reverse the Source Separation Rule as if we are facing CJK extension blocks and doing some regular normalization (which gets rid of compatibility code points). Might be interesting for feeding the LM text from different CJK languages to obtain blocky soup...
  • Idea: We can let the bot generate pseudo-pseudo-Chinese (偽偽中国語, a.k.a. zh@face_white) by feeding it Japanese text without kana and normal Chinese. This normalization may help the LM stir things further by explicitly showing character equivalence, although some extra intervention to revert Han simplification on both scripts can be more helpful.
  • Somehow similar to unifying zh-Hans and zh-Hant with OpenCC, but really more ambitious and playful.