• Add detect_model_family() to identify model family (wan, sdxl, sd, ltx, etc.)
• Add get_pipeline_for_model_family() for proper pipeline selection based on family + task
• Enhance detect_generation_type() to check --image FIRST for I2V detection
• Add support for --image_model, --prompt_image, --prompt_animation as I2V indicators
• Add support for audio/subtitle options as T2V+V2V chaining indicators

This fixes the issue where SDXL models were incorrectly using WanPipeline for I2V tasks, causing type mismatch errors (expected UMT5EncoderModel, got CLIPTextModel). Now SDXL models correctly use DiffusionPipeline or StableDiffusionXLPipeline.