Fuse qkv and fix bart decoding

[gongel]

PR types

New features & Bug fixes

PR changes

APIs

Description

• Fuse qkv for faster decoding.
• Fix: When parallel decoding, the cache size is zero which causes to wrong decoding result.

TODO: Fix other similar models when upgrading to FT 5.x

[paddle-bot]

Thanks for your contribution!

[codecov]

Codecov Report

Merging #5111 (4b74d1d) into develop (9e23230) will increase coverage by 0.06%.
The diff coverage is 0.00%.

@@             Coverage Diff             @@
##           develop    #5111      +/-   ##
===========================================
+ Coverage    49.82%   49.89%   +0.06%     
===========================================
  Files          456      456              
  Lines        65500    65415      -85     
===========================================
+ Hits         32636    32637       +1     
+ Misses       32864    32778      -86     

Impacted Files	Coverage Δ
...lenlp/ops/fast_transformer/transformer/decoding.py	8.25% <0.00%> (+0.53%)	⬆️
...s/fast_transformer/transformer/fast_transformer.py	13.38% <ø> (ø)
paddlenlp/utils/downloader.py	65.92% <0.00%> (+0.44%)	⬆️

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