fix: robust JSON extraction for mixed LLM responses

[SergioChan]

SummarynnHarden backend JSON parsing for LLM responses so mixed outputs (markdown fences, pre/post text) are handled more robustly, reducing 500 errors reported during ontology generation.nn## Changesnn- Updated LLMClient.chat() to remove <think ...>...</think> tags case-insensitivelyn- Added LLMClient._extract_json_payload() to normalize and extract JSON from noisy model responsesn- Updated chat_json() to parse extracted payload instead of raw contentn- Added unit tests in backend/tests/test_llm_client_json_extract.py for fenced JSON and mixed-text extractionnn## Testingnn- Added targeted unit tests for extraction behaviorn- Could not execute tests in this environment because backend dev dependencies are not installed (pytest, flask missing)nnFixes #64