Allow to set Adam beta1, beta2 in TrainingArgs

[gonglinyuan]

In some models, beta1 and beta2 in Adam optimizer are set to be different from the default values (0.9, 0.999). For example, RoBERTa set beta2 = 0.98. It is thereby necessary to add beta1 and beta2 in TrainingArgs if the user wants to fine-tune RoBERTa and other similar models. Also, another hyperparameter of Adam, adam_epsilon, has already been added to TrainingArgs. For the purpose of consistency, it would be better of adam_beta1 and adam_beta2 are also added.

[codecov]

Codecov Report

Merging #5592 into master will increase coverage by 1.21%.
The diff coverage is 75.00%.

@@            Coverage Diff             @@
##           master    #5592      +/-   ##
==========================================
+ Coverage   76.95%   78.16%   +1.21%     
==========================================
  Files         145      145              
  Lines       25317    25319       +2     
==========================================
+ Hits        19482    19790     +308     
+ Misses       5835     5529     -306     

Impacted Files	Coverage Δ
src/transformers/trainer.py	37.96% <0.00%> (ø)
src/transformers/trainer_tf.py	16.53% <ø> (ø)
src/transformers/optimization_tf.py	57.65% <100.00%> (ø)
src/transformers/training_args.py	78.00% <100.00%> (+0.44%)	⬆️
src/transformers/modeling_tf_roberta.py	43.98% <0.00%> (-49.38%)	⬇️
src/transformers/generation_tf_utils.py	83.20% <0.00%> (-1.76%)	⬇️
src/transformers/file_utils.py	79.26% <0.00%> (-0.34%)	⬇️
src/transformers/modeling_tf_mobilebert.py	96.77% <0.00%> (+73.38%)	⬆️

---

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 cfbb982...9bdb1a5. Read the comment docs.

[LysandreJik]

LGTM, love this! @julien-c @jplu

[jplu]

LGTM! Really nice!!!

[julien-c]

I'm fine with this