[Model] Add use_qk_norm option for Cohere model

[tlopex]

This pr updates use_qk_norm option for Cohere series models like Command-R-Plus.

[tlopex]

cc @MasterJH5574
Honestly, I don't know the way that split-and-combine qkv like this is good or not. And since the Command-R-Plus is too large that my device cannot hold it, it will be extremely helpful for me if you can give me some suggestions and test with it.

[MasterJH5574]

@tlopex Sorry for the delayed response. Split-and-combine is okay for now. Could you try the 4b quantized version on your end? If that's not possible I can find a way to test on my side.

[tlopex]

@MasterJH5574
My device is equipped with 16GB of GPU memory, and after applying INT4 quantization, the 55.1B c4ai-command-r-plus-4bit model is estimated to still require around 27.5GB of memory, I assume that it will be difficult for me to test this on my side. So I may need your assistance.

[tlopex]

@MasterJH5574 Hi! Could you find a way to test this, so we can determine if the PR is ready for merging?