Implements tool call handling for FunctionGemma models which use a
unique format: <start_function_call>call:name{key:<escape>val<escape>}<end_function_call>

Changes:
- Add COMMON_CHAT_FORMAT_FUNCTION_GEMMA enum
- Implement grammar builder for FunctionGemma format
- Implement parser for FunctionGemma tool calls
- Add template detection for <start_function_call> markers
- Add google-functiongemma.jinja template
- Add unit tests and e2e test entries

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

Parser: Use state machine to find closing brace, properly handling `}`
characters inside `<escape>` delimiters. Previously used [^}]* regex
which would break on values like `{key:<escape>foo}bar<escape>}`.

Template: Output tool calls in native FunctionGemma format
`{key:<escape>value<escape>}` instead of JSON format `{"key":"value"}`.
This ensures the model sees correct examples in conversation history.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Add common_peg_preserved_parser type that matches preserved tokens
(like <tool_call>, <escape>) at the token level when token info is
available, with graceful fallback to text matching.

New features:
- common_peg_token_span struct for token boundary info
- common_peg_parse_context extended with optional token spans
- preserved(literal, token_id) builder method
- Token-aware matching: if token_id set and context has tokens,
  match by token ID; otherwise fall back to text

This is the infrastructure for token-aware grammars as described in
docs/development/token-aware-grammars.md.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Add comprehensive tests for the preserved parser:
- Text fallback (success, fail, partial)
- Token-aware matching when token info available
- Token mismatch detection
- Sequence with preserved tokens
- Qwen3-Coder-like parser demo (with and without token info)

The tests demonstrate how preserved() enables parsers to match
special tokens at the token level when token boundary info is
available, falling back to text matching otherwise.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

- Rename common_peg_preserved_parser to common_peg_token_parser
- Rename preserved() method to token() for cleaner API
- Token IDs are now resolved at parse time from context's token_ids map
  instead of being hardcoded in the parser definition
- Add token_ids map to common_peg_parse_context for runtime resolution
- Update FunctionGemma parser to use new token() API
- Wire up FunctionGemma PEG parser (COMMON_CHAT_FORMAT_PEG_FUNCTION_GEMMA)
- Update all tests to use new token() method and provide token_ids map

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Add LLAMA_GRETYPE_TOKEN_LITERAL for token-aware grammar matching:
- Matches by token ID when literal resolves to single token
- Falls back to character-by-character matching otherwise

Also adds until_token() to PEG parser and updates to_gbnf() to emit
@"..." syntax for token() primitives, enabling token-aware constrained
decoding in generated grammars.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>