Support qwen2 model, optimize phi3 model, revise model loading strategy

[guoqingbao]

Key changes for #44 :

1. QWen2 model added
2. Performance of Phi3 optimized
3. Redundant transpose removed in decoding stage for paged attention
4. Model loading strategy changed: given a model type, specify the local weight path or concret model id, this enables loading of different models under the same model architecture, for example, llama3 for model type of "llama"
5. ReadMe updated to reflect newly supported models and corresponding usage

Tested case:

cargo run --release -- --port 2000 --weight-path /home/qwen2-1.8b/ qwen2 --repeat-last-n 64

or

cargo run --release -- --port 2000 --model-id Qwen/Qwen1.5-1.8B-Chat qwen2 --repeat-last-n 64

Around 150 tokens/s achieved for qwen2 1.8B on A100 (mixed precision of FP32 and BF16).

[EricLBuehler]

Thank you for adding this, it is great. There are a bunch of unused items warnings when you build this, can you please check that?

[guoqingbao]

Thank you for adding this, it is great. There are a bunch of unused items warnings when you build this, can you please check that?

Fixed

[EricLBuehler]

Thank you!