Add SwiGLU for auto Llama

[From00]

PR types

Performance optimization

PR changes

Models

Description

自动并行Llama模型支持SwiGLU，以优化性能和显存。
Llama1-13B-PP2-SD4-GBS128-ACC32配置，单机A100提速0.9%，相关数据：

关联PR： #7967

遗留TODO： 补齐框架SwiGLU算子切分推导规则（测试配置未开mp，走兜底规则）

[paddle-bot]

Thanks for your contribution!

[codecov]

Codecov Report

Attention: Patch coverage is 40.00000% with 12 lines in your changes are missing coverage. Please review.

Project coverage is 56.42%. Comparing base (dff7dcc) to head (1b113be).

Files	Patch %	Lines
paddlenlp/transformers/llama/modeling_auto.py	40.00%	6 Missing ⚠️
...ddlenlp/transformers/llama/modeling_auto_static.py	40.00%	6 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff            @@
##           develop    #8038   +/-   ##
========================================
  Coverage    56.42%   56.42%           
========================================
  Files          589      589           
  Lines        90252    90266   +14     
========================================
+ Hits         50921    50929    +8     
- Misses       39331    39337    +6     

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

[zhiqiu]

LGTM