Refactor turbomind attention by precomputing rotary embed

[irexyc]

Motivation

Calculate cos/sin in advance and reduce the parameters of the prefill/decode kernel

[lvhan028]

"internlm/internlm2_5-7b-chat-1m" long-context test failed

[lzhangzz]

Do we really need float for RoPE? Most models store them as half or bf16.

[zhulinJulia24]

@irexyc run generation,

CUDA_VISIBLE_DEVICES=4,5 python3 benchmark/profile_generation.py /nvme/qa_test_models/mistralai/Mixtral-8x7B-Instruct-v0.1 --backend pytorch -c 8 256 -ct 128 128 2048 128 -pt 1 128 128 2048 --tp 2 --cache-max-entry-count 0.8

only use 1 gpu and process hangs.