Add support for 336px CLIP vision encoder

[yxjiang]

• Change default vision encoder to clip-vit-large-patch14-336 (336px)
• Add image_size property to CLIPVisionEncoder
• Add num_visual_tokens property that calculates dynamically from image size
• Update max_length defaults to 1024 for both Phase 1 and Phase 2
• Add validation warnings for insufficient max_length
• Update checkpoint filenames to include pixel resolution (e.g., checkpoint_phase1_fp16_336px.pt)
• Update datasets to auto-calculate num_visual_tokens from image processor
• Update training scripts to pass num_visual_tokens explicitly