non recurrent mode for variational dropout

[jsenellart]

variational dropout as described in Gal et al., 2016 does not have the expected result for NMT. Adding a new mode variational_non_recurrent for further exploration.

[codecov-io]

Codecov Report

Merging #390 into master will decrease coverage by <.01%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master     #390      +/-   ##
==========================================
- Coverage   69.62%   69.61%   -0.01%     
==========================================
  Files          74       74              
  Lines        6282     6290       +8     
==========================================
+ Hits         4374     4379       +5     
- Misses       1908     1911       +3

Impacted Files	Coverage Δ
onmt/modules/Encoder.lua	92.2% <100%> (ø)	⬆️
onmt/modules/LSTM.lua	98.3% <100%> (ø)	⬆️
onmt/modules/GRU.lua	94.44% <100%> (ø)	⬆️
onmt/train/Trainer.lua	17.69% <0%> (-0.21%)	⬇️
onmt/data/Preprocessor.lua	92.97% <0%> (+0.05%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update b3b26fc...7cf6909. Read the comment docs.