Performance of Vulkan backend looks amazing

[foldl]

I have just updated chatllm.cpp to use ggml from the last commit (0f2bbe6). Performance of Vulkan backend looks amazing: it is much faster than the CUDA backend (in this test to be precious).

Command line options: -m qwen2.5-1.5b.bin -ngl all -t 0 -p "write a quick sort function in python" --max_length 200

Hardware: 2080TI with 22GB.

Backend	t/s (prompt evel)	t/s (generation)
CUDA	335	77
Vulkan	802	103