Different quantization for K and V cache

[geomi737]

Right now changes in K cache and V cache quantization changes both of them. But the most effective strategy to quantize them is to set K to Q8_0 (Changes in K cache have more chances to add changes to the final output between default model and with K quantinized) and V to Q4_0.