Sanitize dirty JSON tool responses

[dmytryk-here]

Summary

• Sanitize common LLM output artifacts before dirty JSON extraction.
• Handles markdown code fences, XML-like invoke artifacts, function-style prefixes, and leading text before JSON.
• Adds a focused regression test for wrapped tool-call JSON parsing.

Tests

. .venv/bin/activate && pytest tests/test_dirty_json.py tests/test_extract_tools_json_parse.py tests/test_tool_request_normalization.py

Result:

13 passed

## Notes

Searched for overlapping upstream PRs related to json_parse_dirty / sanitizer behavior before preparing this change.