cutlass 3.x gemm on sm90

[ckl117]

PR types

Performance optimization

PR changes

Others

Description

增加cutlass 3.x FP8 GEMM 代码生成和调优；
默认不开启调优，通过环境变量FLAGS_use_cutlass_device_best_config_path=tune控制cutlass FP8 gemm调优功能，如果未设置变量、为空或者defalut则使用默认配置，否则使用指定的json文件配置，增加易用性；
去除nvcc编译选项-O3，该选项或造成部分cutlass 配置失效；

[paddle-bot]

Thanks for your contribution!

[DrownFish19]

安装方式已经变化，setup_cuda中cutalss已经通过submodule添加

[codecov]

Codecov Report

Attention: Patch coverage is 0% with 7 lines in your changes missing coverage. Please review.

Project coverage is 52.94%. Comparing base (4b02477) to head (a4f85e7).

Files with missing lines	Patch %	Lines
...erimental/transformers/fused_transformer_layers.py	0.00%	7 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #9398      +/-   ##
===========================================
+ Coverage    52.93%   52.94%   +0.01%     
===========================================
  Files          688      688              
  Lines       109379   109384       +5     
===========================================
+ Hits         57899    57917      +18     
+ Misses       51480    51467      -13     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[DrownFish19]

PaddleNLP-CI跳过了此处测试，合入需验证版本升级带来影响。