fix: disable json_object response_format for gemma models

[octo-patch]

Fixes #344

Problem

When using LM Studio with gemma models (e.g. gemma-4-e2b), the
visualization and other JSON-structured features fail with:

openai.BadRequestError: Error code: 400 - {'error': "'response_format.type' must be 'json_schema' or 'text'"}

The root cause: the lm_studio binding has supports_response_format: True
in PROVIDER_CAPABILITIES, so the code sends response_format={"type": "json_object"}.
However, newer Gemma models only accept json_schema or text — they do not
support the legacy json_object type.

Solution

Add "supports_response_format": False to the existing "gemma" entry in
MODEL_OVERRIDES inside deeptutor/services/llm/capabilities.py.

When supports_response_format returns False, the code omits the
response_format parameter entirely. The existing extract_json_object
utilities in the visualize and math-animator agents already parse
structured JSON from plain-text responses, so all callers work correctly
without response_format being set.

This is the same pattern already used for DeepSeek and Anthropic models.

Testing

• Direct Python assertions covering gemma-4-e2b, gemma-3-4b, gemma-2-9b all return False
• Non-gemma LM Studio models (mistral-7b, llama-3) continue to return True
• Added test_gemma_response_format_disabled to tests/services/llm/test_capabilities.py

[mr-Lime197]

this problem persists with qwen-like architectures.

[pancacake]

Thanks for your contribution!