This commit introduces several enhancements to model handling, chat functionality, and the user interface. Key changes include refining the model store, improving the chat experience with auto-scroll and auto-restore, and increasing the default token limit for GGUF models.

### Model Management & Store

-   **Curated Model Lists:** The built-in model lists for both NPU and CPU have been refined by removing the `CuteYukiMix` model.
-   **GGUF Repository Cleanup:** The default list of GGUF model repositories has been significantly trimmed down to a more curated selection of three general-purpose models.
-   **Improved GGUF File Detection:** The logic for discovering GGUF files in repositories is now more robust. It actively filters out auxiliary files like `mmproj`, `vision-adapter`, and `projector` to prevent them from being listed as standalone models.

### Chat Experience

-   **Auto-Restore Last Session:** The app now remembers and automatically restores the last active chat and the last used model on startup, allowing for a seamless continuation.
-   **Smarter Auto-Scroll:** The chat view's auto-scrolling is now more intelligent. It automatically stops if the user scrolls up manually and resumes once they scroll back to the bottom.
-   **Increased Generation Limit:** The default `max_tokens` for GGUF models has been increased from 512 to 2048, enabling longer and more detailed responses.
-   **Dynamic `max_tokens`:** The `sendChat` function now dynamically reads the `max_tokens` value from the currently loaded model's configuration, falling back to 2048 if not specified.
-   **System Prompt Integration:** The chat now automatically includes the system prompt defined in a model's configuration, ensuring models behave as intended.
-   **More Reliable Stop:** The process for stopping a text generation has been improved with a small delay to ensure final metrics are received before the job is canceled.

### UI/UX Refinements

-   **Redesigned `ModelCard`:** The model cards in the Model Store have been redesigned to be more compact and modern, using text-based tags and chips instead of `AssistChip` components.
-   **Grouped Model Store:** Models in the store are now grouped by their source repository (for GGUF) or hardware target (CPU/NPU), making the list easier to navigate.
-   **Theme Enhancements:** Updated the app theme for better Day/Night support and transparent system bars.
-   **Minor Formatting:** Corrected spacing in file size strings (e.g., "1.1GB" to "1.1 GB").

…haul