clip : refactor, add image_manipulation and llava_uhd classes

[ngxson]

This PR partially resolves #12944 , but the public API for counting tokens will be added in a follow-up PR (cc @mattjcly )

The current PR only contains internal changes, the public API is not touched.

Motivation

The idea of this refactoring comes from my blog post where I documented the preprocessing step of vision models. This step is purely algorithmic. The "slicing" and "grid" systems are somewhat equivalent to pre-tokenizer regex in text models.

Currently, only some vision arch using this system:

• minicpm-v: grid is dynamically calculated
• llava-1.6: grid is decided using a list of grid pinpoints

Implementation

After this refactoring, the overall flow is:

• Given an image, get_slice_instructions is called to decide how to slice the image
• After getting the instruction, we use image_manipulation class to actually modify the image (crop, resize, pad) ; The output of this step are slices
• Convert slices to clip_image_f32 and normalize them

sequenceDiagram
    participant User
    participant clip_image_preprocess
    participant llava_uhd as get_slice_instructions
    participant slice_image as slice_image
    participant image_manipulation
    participant normalize as normalize_image_u8_to_f32
    
    User->>clip_image_preprocess: submit clip_image_u8
    clip_image_preprocess->>llava_uhd: request slicing instructions
    llava_uhd-->>clip_image_preprocess: return list of slice_instructions
    
    clip_image_preprocess->>slice_image: call with slice_instructions
    
    loop for each slice_instruction
        slice_image->>image_manipulation: crop, resize, pad image
        image_manipulation-->>slice_image: return processed image slice
    end
    
    slice_image-->>clip_image_preprocess: return std::vector<clip_image_u8_ptr>
    
    loop for each clip_image_u8_ptr
        clip_image_preprocess->>normalize: convert image
        normalize-->>clip_image_preprocess: return clip_image_f32
    end
    
    clip_image_preprocess-->>User: return list of clip_image_f32

Loading

Test

Test script confirms that this does not break anything:

OK:   llama-gemma3-cli ggml-org/gemma-3-4b-it-GGUF:Q4_K_M
OK:   llama-llava-cli cmp-nct/Yi-VL-6B-GGUF:Q5_K
OK:   llama-llava-cli guinmoon/MobileVLM-3B-GGUF:Q4_K_M
OK:   llama-llava-cli THUDM/glm-edge-v-5b-gguf:Q4_K_M
OK:   llama-llava-cli second-state/Llava-v1.5-7B-GGUF:Q2_K
OK:   llama-llava-cli cjpais/llava-1.6-mistral-7b-gguf:Q3_K
OK:   llama-llava-cli ibm-research/granite-vision-3.2-2b-GGUF:Q4_K_M
OK:   llama-minicpmv-cli second-state/MiniCPM-Llama3-V-2_5-GGUF:Q2_K
OK:   llama-minicpmv-cli openbmb/MiniCPM-V-2_6-gguf:Q2_K
OK:   llama-minicpmv-cli openbmb/MiniCPM-o-2_6-gguf:Q4_0
OK:   llama-qwen2vl-cli bartowski/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M