mtmd: add --image-warmup-tokens

[SmartestWashingMachine]

Noticed that a Q4_1 Qwen3-VL 2B mmproj reserved a surprisingly large amount of memory for one of its compute buffers during the image warmup part - larger than the model itself!

Looking at the code, it seems that image warmup sizes are hard-coded for different models. For Qwen3-VL it's 2116 tokens, or an image with 1472 x 1472 dimensions. If I understand correctly, llama-cpp initially reserves memory proportional to the size of that warmup image.

But some users may be certain that their images will never exceed some dimensions (e.g: OCR'ing single lines of text or their preprocessing pipeline caps images at 512 x 512), so they may want a smaller max image warmup size, reducing the memory consumption.

A few hundred MB cut down doesn't sound like much on its own, but it might help for development on the edge.

Before (Initial behavior)

With --image-warmup-tokens 256

[ngxson]

I'd prefer having a warmup option like in libllama. image-warmup-tokens is quite low-level I think, it probably not very future-proof as we can probably have other strategy for warming up in the future.

The warmup option should matches the common_params::warmup

[ngxson]

Superseded by #17652

It will be more suitable in your use case, as you know the image size in advanced, not the number of tokens

[SmartestWashingMachine]

Oh, that's even better. Thanks for taking the time to look into this!