Add on hold for "loading checkpoint or unet or so on" only start when running their own function to save VRAM

### Feature Idea

I use LLM node to create prompt inside Comfyui, however when loading LLM checkpoint, ComfyUI also loading Flux checkpoint which increate VRAM alot when I only generate prompt then when ComfyUI begin generate image, ComfyUI does not unload Vram of LLM checkpoint.

### Existing Solutions

Make checkpoint loading node wait and begin loading only when receive some signal like prompt or logic value or some else. Release Vram of checkpoint whenever its finish generate image (for example: SAM model, after generate mask image unload VRAM, LLM or VLM node after creating prompt then unload model and cache and only load checkpoint into VRAM only when receive input either image or text)

### Other

_No response_