Add FramePack sampler for FramePack-F1

[kabachuha]

This Pull request adds support for the new Framepack variant Framepack F1 (forward sampling) in a form of a new FramePackSamplerF1 node.

Here you can see a test result:

FramePack_00001.mp4

---

P.S. Converted Framepack F1 to kijai format for those who like to work with single files:

https://huggingface.co/kabachuha/FramePack_F1_I2V_HY_20250503_comfy/tree/main

[neph1]

Tested, and it works great!

Edit: What method did you use to create the fp8 file? I've looked around but haven't found the process documented.

[kabachuha]

@neph1 I put this code (derived from the native format loading) inside the Download Framepack node:

    state_dict = transformer.state_dict()
    from safetensors.torch import save_file
    
    save_file(state_dict, 'FramePack_F1_I2V_HY_20250503.safetensors')
    with open('keys.txt', 'w', encoding='utf-8') as f:
        for key in list(state_dict.keys()):
            f.write(f"{key}\n")
    
    params_to_keep = {"norm", "bias", "time_in", "vector_in", 
                        "guidance_in", "txt_in", "img_in"}
    
    # Convert parameters
    converted_dict = {}
    for key, tensor in state_dict.items():
        if any(pattern in key for pattern in params_to_keep):
            # Keep original dtype for specified parameters
            converted_dict[key] = tensor
        else:
            # Convert to float8_e4m3fn for other parameters
            if tensor.is_floating_point():
                converted_dict[key] = tensor.to(torch.float8_e4m3fn)
            else:
                converted_dict[key] = tensor
    
    # Save the converted model
    save_file(converted_dict, "FramePack_F1_I2V_HY_20250503_fp8_e4m3fn.safetensors")
    print("Successfully saved converted model")
    
    raise 'debug'
    ```

[kabachuha]

@kijai are you ready to review this?

[kajfblsdkgnlsndg]

When can this update be completed approximately

[kajfblsdkgnlsndg]

After testing for a period of time, I felt that the F1 model had some flaws. Every second of the transition screen would experience shaking and unnatural transitions. I tried using the F1 sampler and the original FramePack model, which worked quite well.

[wardboumans]

Can you add End image latent?