feat(utils): add json_utils module for robust LLM JSON parsing

[JasonOA888]

Problem

Some LLM models don't respect json_object format and return markdown-wrapped JSON:

```json
{"key": "value"}

This causes `json.loads()` to fail with JSONDecodeError.

Affected models: MiniMax M2.5, GLM-4.7, GLM-5

Related issues: #72, #64, #58, #48

## Solution

Add `json_utils.py` module with:

- `clean_llm_json_response()` - Strip markdown code blocks
- `parse_llm_json()` - Parse with auto-cleanup, optional default value
- `safe_parse_llm_json()` - Non-throwing version returning `(success, result)`

## Usage

```python
from app.utils import parse_llm_json

# Auto-strips ```json ... ``` wrappers
result = parse_llm_json(llm_response)

# With default fallback
result = parse_llm_json(llm_response, default={})

# Safe version (no exceptions)
success, result = safe_parse_llm_json(llm_response)

Changes

• New file: backend/app/utils/json_utils.py
• Update: backend/app/utils/llm_client.py uses new helper
• Update: backend/app/utils/__init__.py exports new functions