Fix GPT OSS tool call parser and test configuration
- Move TOOL_OPEN tag to wrap only "to=functions." literal instead of
the entire assistant prefix, matching the expected streaming behavior
- Add missing p.tag(Tag::TOOL, ...) wrapper to tool-role pattern
- Fix content type matching: use p.until("<|message|>") to match any
content type (e.g., " json") without requiring <|constrain|> token
- Set SupportsReasoningOnly::No since template always outputs final
channel content (outputs literal "None" when content is null)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
