Optimize ft when using padded vocabulary

[FrostML]

PR types

Function optimization

PR changes

Others

Description

• Optimize FT when using padded vocabulary
• Fix fp16 usage in eager mode
• Fix UNIMO doc
• Fix UnifiedTransformer original generation
• Fix UNIMO origin generation
• Fix MBART cannot generate with use_cache=False
• Fix BART cannot generate with use_cache=False
• Fix T5 cannot generate with use_cache=False

[guoshengCS]

是否有不依赖fluid的接口来判断，另外写上注释说明需要这个分支的原因

[FrostML]

必须得有 paddle.fluid.framework，paddle.framework 没有暴露这个变量。

[codecov]

Codecov Report

Merging #3936 (69b2978) into develop (5cc4b21) will increase coverage by 0.00%.
The diff coverage is 51.35%.

@@           Coverage Diff            @@
##           develop    #3936   +/-   ##
========================================
  Coverage    36.30%   36.30%           
========================================
  Files          419      419           
  Lines        59222    59254   +32     
========================================
+ Hits         21498    21514   +16     
- Misses       37724    37740   +16     

Impacted Files	Coverage Δ
...lenlp/ops/fast_transformer/transformer/decoding.py	7.47% <0.00%> (-0.03%)	⬇️
paddlenlp/transformers/bart/modeling.py	84.16% <44.44%> (-1.15%)	⬇️
paddlenlp/transformers/mbart/modeling.py	80.73% <44.44%> (-1.03%)	⬇️
paddlenlp/transformers/t5/modeling.py	85.18% <77.77%> (-0.11%)	⬇️
...lenlp/transformers/unified_transformer/modeling.py	82.06% <100.00%> (ø)
paddlenlp/transformers/unimo/modeling.py	82.58% <100.00%> (ø)
paddlenlp/transformers/roberta/modeling.py	90.21% <0.00%> (+0.36%)	⬆️

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.