refactored beam search according to torch implementation

[patrickvonplaten]

1:1 translation of PR #3135 from Pytorch to TF.

[patrickvonplaten]

good to merge for me

[codecov-io]

Codecov Report

Merging #3148 into master will decrease coverage by 1.04%.
The diff coverage is 95.83%.

@@            Coverage Diff             @@
##           master    #3148      +/-   ##
==========================================
- Coverage   78.03%   76.98%   -1.05%     
==========================================
  Files          98       98              
  Lines       16573    16583      +10     
==========================================
- Hits        12932    12766     -166     
- Misses       3641     3817     +176

Impacted Files	Coverage Δ
src/transformers/modeling_tf_utils.py	91.47% <95.83%> (-1.95%)	⬇️
src/transformers/modeling_tf_pytorch_utils.py	8.72% <0%> (-81.21%)	⬇️
src/transformers/modeling_roberta.py	85.71% <0%> (-10%)	⬇️
src/transformers/modeling_xlnet.py	73.48% <0%> (-2.3%)	⬇️
src/transformers/modeling_ctrl.py	96% <0%> (-2.23%)	⬇️
src/transformers/modeling_openai.py	80.2% <0%> (-1.35%)	⬇️
src/transformers/modeling_utils.py	94.56% <0%> (-0.16%)	⬇️

---

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 0001d05...2861c9d. Read the comment docs.

[LysandreJik]

LGTM

[thomwolf]

Good to me as well, merging