Fix TF optimizer

[jplu]

Fix #6560. Now the parameter experimental_aggregate_gradients should be correctly taken into account.

[codecov]

Codecov Report

Merging #6717 into master will decrease coverage by 0.33%.
The diff coverage is 50.00%.

@@            Coverage Diff             @@
##           master    #6717      +/-   ##
==========================================
- Coverage   79.77%   79.44%   -0.34%     
==========================================
  Files         156      156              
  Lines       28392    28392              
==========================================
- Hits        22650    22556      -94     
- Misses       5742     5836      +94     

Impacted Files	Coverage Δ
src/transformers/optimization_tf.py	57.65% <50.00%> (ø)
src/transformers/modeling_tf_xlnet.py	21.12% <0.00%> (-71.05%)	⬇️
src/transformers/modeling_roberta.py	77.37% <0.00%> (-19.71%)	⬇️
src/transformers/modeling_tf_utils.py	86.64% <0.00%> (-0.66%)	⬇️
src/transformers/file_utils.py	82.41% <0.00%> (+0.25%)	⬆️
src/transformers/generation_tf_utils.py	86.46% <0.00%> (+1.25%)	⬆️
src/transformers/tokenization_xlnet.py	90.09% <0.00%> (+23.42%)	⬆️
src/transformers/tokenization_mbart.py	95.31% <0.00%> (+39.06%)	⬆️
src/transformers/modeling_tf_albert.py	90.90% <0.00%> (+69.43%)	⬆️

---

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 841f071...5281631. Read the comment docs.

[LysandreJik]

Great, LGTM!

[sgugger]

Nice and short, thanks!