Reject malformed JSON in LLM coercion

[Fu Xiaonan (FU-max-boop)]

Summary

• accept an empty JSON object only when the complete LLM output is {} (allowing surrounding/interior whitespace)
• reject malformed or mixed JSON that extract_json reports with its {} failure sentinel
• add regression coverage for the reported mismatched-brace response and related malformed outputs

Validation

• pytest tests/unit/test_llm_chat.py -k coerce_to_dict_or_list -q — 2 passed
• pytest tests/unit/test_llm_chat.py -k coerce -q — 12 passed
• pytest tests/unit/test_utils.py -k 'extract_json or name_or_empty or repeat_on_error' -q — 3 passed\n- black --check tests/unit/test_llm_chat.py --target-version py310\n- python -m py_compile tinytroupe/utils/llm.py tests/unit/test_llm_chat.py\n- git diff --check\n\nFixes utils.llm.LLMChat._coerce_to_dict_or_list allows misformatted JSONs through #159