Add DeepSeek and Qwen support with enhanced AI mistake analysis

[xhuandy666]

Overview

This PR upgrades LeetCode EasyRepeat from v1.0.6 to v1.0.9 with first-class support for major Chinese LLM providers and a more reliable, useful AI-assisted mistake-analysis workflow.

Key changes

Chinese and multi-provider model support

• Add official DeepSeek Cloud support, including API-key validation, dynamic model discovery, exact model selection, the extension-side gateway, and the Python Auto-Fix backend.
• Add Qwen support through Alibaba Cloud DashScope's OpenAI-compatible API. When a gateway does not expose /models, the UI falls back to recommended Qwen models.
• Add a generic OpenAI-compatible provider with a configurable Base URL and custom model ID.
• Preserve the existing Gemini, OpenAI, Anthropic, Ollama, and local deepseek-r1 integrations.
• Honor the provider and model selected by the user without automatic cross-provider failover.

Higher-quality mistake analysis

• Match the analysis language to the extension UI language: Chinese UI produces Chinese analysis, and English UI produces English analysis.
• Include the programming language, failing input, actual output, and expected output in the analysis prompt.
• Ask the model to explain why the submission is wrong, provide the correct approach and corrected code, identify what is still missing compared with a correct solution, and recommend the next step.
• Give constructive starting guidance for empty or severely incomplete submissions without inventing a specific bug.

More reliable submission handling

• Snapshot the editor code at submit time so later editor changes cannot alter the analyzed submission.
• Distinguish code-capture failure from a genuinely empty submission, surface the failure explicitly, and skip the model request when no trustworthy code was captured.
• Avoid writing empty or severely incomplete attempts into Safe Observer or the similar-error store.
• Allow an explicitly configured custom model ID when an OpenAI-compatible gateway does not implement model discovery.

Permissions and backend integration

• Add predefined host access for DeepSeek and DashScope.
• Request Chrome host permission at runtime for user-configured compatible endpoints.
• Extend the Python provider factory and model endpoint for DeepSeek, Qwen, and custom OpenAI-compatible providers.
• Add the missing Anthropic backend dependency and align provider/model endpoint tests with the expanded provider set.

Validation

• Frontend targeted tests: 44/44 passed.
• Python backend tests: 60/60 passed.
• Production Vite build: passed.
• The full Jest run has the same 11 pre-existing failing suites as upstream main; this branch introduces no new full-suite failures.

Testing note

No real paid-provider request was made because no personal API key was used. Provider requests, model discovery, prompt construction, capture states, and backend routing are covered by automated tests.