Fix LFM2 tool calling with nested parentheses in arguments

[tpae]

Summary

• Fix PythonicToolCallParser regex that truncated tool call arguments containing nested parentheses (e.g., requests.get('url') inside a code= argument). The root cause was optional brackets \[?...\]? in the regex -- the Python reference uses required brackets \[...\], which forces the non-greedy .*? to backtrack past inner ) characters. A string-based fallback handles the without-brackets case.
• Add "*.jinja" to downloadModel file patterns so chat_template.jinja is cached alongside model weights for offline robustness.

Context

LFM2-24B-A2B (LiquidAI/LFM2-24B-A2B) uses Pythonic tool call syntax wrapped in <|tool_call_start|> / <|tool_call_end|> tags. When argument values contain nested parentheses (common with code-execution tools like sandbox_run_script), the regex matched the first ) inside the value rather than the correct outer one, producing mangled arguments.

Reference: mlx-lm/tool_parsers/pythonic.py

Test plan

• New test: nested parentheses with brackets ([run_script(code="requests.get('url')")])
• New test: nested parentheses without brackets (run_script(code="print('hello')"))
• All 37 existing ToolTests pass (no regressions)

Checklist

Put an x in the boxes that apply.

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)

[tpae]

@davidkoski any chance this could be merged?

[davidkoski]

Looks good. I wonder if this will be tripped up with batch tokens where it might have parens in trailing text?