Add cosine scheduler, datacollator, update clip

[w5688414]

…rWholeWordMask, Update CLIPVisionTransformer

PR types

• New features

PR changes

• APIs

Description

• DataCollatorForLanguageModeling
• DataCollatorForWholeWordMask
• forward_pre & forward_post
• get_polynomial_decay_schedule_with_warmup

[paddle-bot]

Thanks for your contribution!

[sijunhe]

collator这块的逻辑较为复杂，参考https://github.com/huggingface/transformers/blob/main/tests/trainer/test_data_collator.py 补上单测吧

[w5688414]

collator这块的逻辑较为复杂，参考https://github.com/huggingface/transformers/blob/main/tests/trainer/test_data_collator.py 补上单测吧

已添加

[codecov]

Codecov Report

Merging #4826 (49584b0) into develop (3ae3ea5) will decrease coverage by 0.17%.
The diff coverage is 12.17%.

@@             Coverage Diff             @@
##           develop    #4826      +/-   ##
===========================================
- Coverage    44.52%   44.36%   -0.17%     
===========================================
  Files          445      446       +1     
  Lines        64042    64332     +290     
===========================================
+ Hits         28515    28540      +25     
- Misses       35527    35792     +265     

Impacted Files	Coverage Δ
paddlenlp/data/data_collator.py	19.32% <11.25%> (-14.89%)	⬇️
paddlenlp/trainer/trainer_utils.py	40.67% <12.50%> (-1.34%)	⬇️
paddlenlp/transformers/clip/modeling.py	57.92% <16.66%> (-0.96%)	⬇️
paddlenlp/transformers/bert/tokenizer.py	95.53% <66.66%> (ø)
paddlenlp/utils/downloader.py	64.60% <0.00%> (-0.89%)	⬇️
paddlenlp/transformers/mt5/modeling.py	84.67% <0.00%> (ø)
...dlenlp/experimental/autonlp/text_classification.py	98.30% <0.00%> (ø)
paddlenlp/transformers/mt5/converter.py	0.00% <0.00%> (ø)

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

[sijunhe]

PR没问题，就一个小改动，tests/trainer/test_data_collator.py应该放到tests/data/test_data_collator.py, 和paddlenlp以下对应