What problem does this PR solve?

This PR solves the problem of duplicated and inconsistent retry logic across LLM model implementations.

Background:

Previously, each LLM model class (embedding, CV, rerank, TTS, OCR, speech2text) had its own ad-hoc retry implementation, leading to:

• Duplicated code across 10+ files
• Inconsistent retry behavior between different model providers
• Difficult maintenance when changing retry settings
• Risk of bugs when retry logic wasn't applied uniformly

Solution:

• Extract retry logic into a centralized rag/llm/retry.py module with:
  • @ retry decorator for sync methods (retries on transient errors, re-raises on failure)
  • @ retry_or_fallback decorator for methods that return error tuples
  • is_retryable() - determines if an error is transient (rate limits, server errors)
  • classify_error() - categorizes exceptions into LLMErrorCode for detailed logging
  • async_handle_exception() - shared async error handler
• Apply decorators consistently across all embedding, CV, rerank, TTS, OCR, and speech2text models
• Add comprehensive unit tests (76 tests)

Behavior preserved:

• Same default retry settings (max_retries=5, base_delay=2.0)
• Same retryable error signals (rate limits, 5xx server errors)
• Same backoff jitter formula (base_delay * uniform(10, 150))

Type of change

• New Feature (non-breaking change which adds functionality)
• Refactoring