commit f794b978bf1f6d8a96e2be3697df984666ec44f5
Author: Haoxiang Fei <feihaoxiang@idea.edu.cn>
Date:   

    perf: read 16-bit trie entries in one load and cast enum ordinals
    
    On wasm, wasm-gc and JS, General_Category was 3-8 times as slow as on
    native. Two causes, both in generated code:
    
    - write_value_trie stored 16-bit mid entries (and wide leaf values) as two
      bytes, so each took two bounds-checked loads. Two entries are now packed
      per Int in a ReadOnlyArray[Int], the same data size; one load and a
      shift. ReadOnlyArray[UInt16] would need no shift, but moonc builds it on
      wasm-gc with one array.set per element.
    - The ordinal-to-enum converters were identity matches, which moonc
      compiles to a br_table on wasm and wasm-gc and a switch on JS; only clang
      folds them away. They are now "%identity", as a constructor without
      arguments is its ordinal on every backend. write_enum loses its fallback
      and the generators check that every ordinal names a constructor.
    
    All value tries are regenerated, including the casefold tables.
    
    Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Changes