feat: add thinking mode toggle for local Qwen models

[jalvarezz13]

Summary

• Adds a UI toggle to enable/disable thinking mode for local Qwen3/3.5 models
• When disabled, appends /no_think to user messages so the model skips the <think> reasoning phase
• Significantly improves inference speed on consumer hardware by avoiding unnecessary reasoning token generation
• Toggle only appears when a model with thinking support is selected

Changes

• Model registry: Added supportsThinking: true flag to all Qwen3 and Qwen3.5 local models
• Settings store: New localThinkingEnabled boolean setting (default: false) persisted in localStorage
• Inference pipeline: When thinking is disabled, /no_think is appended to user messages in both non-streaming (modelManagerBridge.js) and streaming (ReasoningService.ts) paths
• UI: Compact toggle row in the local model section of ReasoningModelSelector, visible only when a thinking-capable model is selected
• i18n: Translations added for all 10 supported languages

How it works

Qwen3/3.5 models support hybrid thinking via the /no_think message suffix. When the user disables thinking mode:

1. Non-streaming: modelManagerBridge.runInference() appends /no_think to the user message before sending to llama.cpp
2. Streaming: ReasoningService.processTextStreaming() appends /no_think to the last user message
3. The existing <think> block stripping in localReasoningBridge.js and ReasoningService.ts remains as a safety net

Testing

• npm run build passes
• No new LSP diagnostics introduced
• Tested locally with Electron app running

Closes #512