Improve Gemini and Claude AI feedback parsing and model refresh flow

[yczhangcc]

Description:
Gemini and Claude responses do not always match the same JSON response shape as ChatGPT/OpenAI. This can cause AI feedback parsing failures even when the provider API key and model are valid.

Problems:

• OpenAI, Gemini, and Claude return model output in different response structures.
• Gemini may include response parts, thought parts, or JSON text that needs to be extracted differently.
• Claude returns content blocks instead of the OpenAI-style message content.
• Some model responses include valid feedback wrapped in provider-specific formats.
• The model selection UI says Fetch Models, but this action is better described as refreshing the current available model list.

Goal:
Make AI feedback generation more reliable across providers:

• Parse OpenAI, Gemini, and Claude responses according to each provider’s response format.
• Keep feedback output normalized into the same internal JSON structure.
• Improve error handling when a provider returns invalid JSON.
• Rename Fetch Models to Refresh Models in model selection UI.
• Keep manual model refresh as a fallback when provider model lists change.

Expected result:
Gemini and Claude feedback should display successfully when they return valid feedback content, and model selection should use clearer Refresh Models wording.