llama.cpp

model: correctly support input vision for deepseek4

#28154

Merged