[FEATURE] [WIP] Use softmax with length in attention cells

[ptrendx]

Description

MXNet added support for softmax with length parameter (apache/mxnet#15169) and this PR attempts to use it in attention cells. Work done by me and @blchu.

@eric-haibin-lin Could you help in making sure this works for all models (we tested just BERT and Transformer decoder)?

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

[codecov]

Codecov Report

❗ No coverage uploaded for pull request head (pr_softmax_with_length@c7ae5b4). Click here to learn what that means.
The diff coverage is n/a.

[codecov]

Codecov Report

Merging #910 into master will decrease coverage by 61.26%.
The diff coverage is 7.69%.

@@             Coverage Diff             @@
##           master     #910       +/-   ##
===========================================
- Coverage   90.48%   29.21%   -61.27%     
===========================================
  Files          66       66               
  Lines        6400     6380       -20     
===========================================
- Hits         5791     1864     -3927     
- Misses        609     4516     +3907

Impacted Files	Coverage Δ
src/gluonnlp/model/transformer.py	14.5% <0%> (-76.71%)	⬇️
src/gluonnlp/model/attention_cell.py	21.64% <25%> (-72.99%)	⬇️
src/gluonnlp/model/bilm_encoder.py	15.25% <0%> (-84.75%)	⬇️
src/gluonnlp/model/train/language_model.py	16.47% <0%> (-80.69%)	⬇️
src/gluonnlp/data/batchify/embedding.py	17.96% <0%> (-79.69%)	⬇️
src/gluonnlp/model/sequence_sampler.py	12.11% <0%> (-79.59%)	⬇️
src/gluonnlp/data/sampler.py	18.59% <0%> (-77.89%)	⬇️
src/gluonnlp/data/dataset.py	22.22% <0%> (-76.99%)	⬇️
src/gluonnlp/model/lstmpcellwithclip.py	23.07% <0%> (-76.93%)	⬇️
src/gluonnlp/model/language_model.py	23.25% <0%> (-76.75%)	⬇️
... and 48 more

[eric-haibin-lin]

moved to #1091 for BERT