chore: Use json-repair for json_tools tool-call parsing

[isaac-cf-wong]

Summary

The json_tools parser now uses json_repair.loads() instead of strict JSON parsing so common LLM mistakes (trailing commas, single-quoted strings, etc.) still produce usable tool calls.

Motivation

Chat templates that emit a JSON object inside <tool_call> often get invalid JSON from the model (I encountered this issue frequently when using mlx-community/Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-4bit). json.loads fails on those outputs; json-repair is aimed at repairing that kind of text.

Changes

• mlx_lm/tool_parsers/json_tools.py: parse_tool_call parses the stripped payload with json_repair.loads.
• setup.py: Declare dependency json-repair>=0.58.7 (latest version).
• tests/test_tool_parsing.py: Add test_json_tools_repairs_malformed_json covering trailing commas and single-quoted JSON for the multiply and temperature fixtures. Existing test_parsers cases still assert valid JSON for json_tools.

Tradeoffs

Parsing is more lenient than before: some malformed text may be repaired into a JSON object instead of raising, which could occasionally yield unexpected name / arguments. That is usually acceptable for tool-call extraction but is a behavioral change from strict json.loads.

Testing

python -m unittest tests.test_tool_parsing