Support VLM Eagle3 Training

[RuBing-Yang]

TODOs:

• 重构Trainer/DatasetBuilder/TargetWrapper仓库，根据online/offline LLM/VLM创建
• Qwen3-VL HF Online训练
• Qwen3-VL HF Offline训练
• 支持mrope

AI Summary：

This pull request introduces support for Vision-Language Model (VLM) data handling and refactors trainer and data collator imports for better modularity. The main changes include adding new VLM-specific data collators, updating chat template types to support VLM, introducing new model configuration files, and cleaning up unused imports related to trainers and data collators.

Vision-Language Model (VLM) Support:

• Added VLMDataCollatorWithPadding and supporting tensor padding functions to data_utils.py for handling VLM-specific batch data, including images and videos. [1] [2] [3]
• Updated chat template handling to include a new QWEN3_VL type and its mapping, enabling VLM-specific system prompts and headers. [1] [2]
• Added new model configuration files for VLM models: qwen3-vl-4b-eagle3.json and qwen3-vl-4b-eagle3-mrope.json. [1] [2]

Trainer and Data Collator Refactoring:

• Removed unused imports and references to OfflineEagle3Trainer, OnlineEagle3Trainer, and DataCollatorWithPadding in __init__.py files, replacing them with Eagle3TrainerFactory for cleaner and more maintainable code. [1] [2] [3] [4] [5]

Dataset Builder Modularization:

• Added a new dataset_builder module with factory and builder classes for both online and offline LLM/VLM datasets, improving extensibility and code organization.