This commit introduces comprehensive support for using different backbone
models beyond LLaVA 1.5 13B, including:

Model Registry (safe/models/model_registry.py):

• Centralized registry for all supported models
• ModelSpec dataclass with architecture details
• Support for Vision-Language, Text-Only, and API model types
• Pre-registered models: LLaVA 1.5 (7B/13B), BLIP2, Llama 3.1/3.2,
  Mistral 7B, Qwen 2.5, Phi-3, Gemma 2, and API models (Gemini, GPT-4)

Base Model Abstraction (safe/models/base_model.py):

• Abstract BaseModel class with unified interface
• TextOnlyModel for Llama, Mistral, Qwen, Phi, Gemma, etc.
• VisionLanguageModel for LLaVA, BLIP2
• APIModel for Gemini Flash/Pro, GPT-4o, Claude
• Factory function create_base_model() for easy instantiation
• Support for Flash Attention 2 and quantization (8-bit, 4-bit)

Updated BaseVLModel (safe/models/base_vl.py):

• Automatic model type detection from model name
• Support for text-only models without vision encoder
• New properties: has_vision, hidden_size, num_layers, num_attention_heads
• get_decoder_layers() method for layer hook injection
• Flash attention and quantization support for text-only models

Model Configurations (configs/model_configs.py):

• Added 15+ new model configurations:
  • LLaVA 1.5 7B (smaller variant)
  • Llama 3.1 8B / 8B-Instruct
  • Llama 3.2 3B / 1B (small efficient models)
  • Mistral 7B v0.3
  • Qwen 2.5 7B
  • Phi-3 Mini 4K
  • Gemma 2 9B / 2B
  • Gemini Flash / Pro (API)
  • GPT-4o / GPT-4o Mini (API)
• Each config includes optimized fusion layer indices and settings
• Convenient aliases: llama-8b, llama-3b, llama-1b

This enables experimenting with modern LLMs like Llama 3.1 and smaller
efficient models for faster iteration and reduced memory requirements.