peg-parser: add unicode-aware trie for GBNF exclusion patterns
Fixes GBNF grammar generation for templates with Unicode tokens.
DeepSeek R1/V3 use fullwidth vertical line (U+FF5C) and lower one
eighth block (U+2581) in their special tokens. The byte-based trie
was generating invalid UTF-8 prefixes in exclusion patterns.
Changes:
- Add unicode_trie that works with code points instead of bytes
- Update gbnf_escape_codepoint to handle non-ASCII characters
- Add tests for DeepSeek-style Unicode token formats
Enables: deepseek_r1:experimental, deepseek_v3_1:experimental
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>