Retry Mechanism to Handle Different Error Types

[ChenZiHong-Gavin]

I've noticed that the current implementation assumes that all models return the same error types as OpenAI, specifically RateLimitError, APIConnectionError, and Timeout. This assumption might not hold true for all models, especially when integrating with different APIs or services.
for example, ollama model:

If so, I may submit a PR to fix this.