[BUGFIX] avoid using dict for attention cell parameter creation

[eric-haibin-lin]

Description

To circumvent apache/mxnet#17056 which causes bert distributed training to diverge

Checklist

Essentials

• PR's title starts with a category (e.g. [BUGFIX], [MODEL], [TUTORIAL], [FEATURE], [DOC], etc)
• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage
• Code is well-documented

Changes

• Feature1, tests, (and when applicable, API doc)
• Feature2, tests, (and when applicable, API doc)

Comments

• If this change is a backward incompatible change, why must this change be made.
• Interesting edge cases to note here

cc @dmlc/gluon-nlp-team

[codecov]

Codecov Report

Merging #1050 into master will increase coverage by 0.03%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master    #1050      +/-   ##
==========================================
+ Coverage   88.24%   88.27%   +0.03%     
==========================================
  Files          67       67              
  Lines        6252     6252              
==========================================
+ Hits         5517     5519       +2     
+ Misses        735      733       -2

Impacted Files	Coverage Δ
src/gluonnlp/model/attention_cell.py	96.22% <100%> (+0.62%)	⬆️
src/gluonnlp/model/transformer.py	91.63% <0%> (+0.32%)	⬆️

[mli]

Job PR-1050/1 is complete.
Docs are uploaded to http://gluon-nlp-staging.s3-accelerate.dualstack.amazonaws.com/PR-1050/1/index.html

[leezu]

Would the non-deterministic parameter order only happen with Python 3.5 and lower? Or did it happen on Python 3.6+?