Add tokenizer to Trainer

[sgugger]

Not entirely sure about this change as there is a trade-off API complexity/ease of use.

This PR adds tokenizer as an optional argument to Trainer (if this is approved, will do the same for TFTrainer, I have a few recent changes to port there but was mainly waiting for @jplu to be back from vacation to make the two APIs on par).

The benefit is that:

• we can have a smart default data_collator that will automatically pad examples if the tokenizer is provided, so the user doesn't have to learn about data_collators for simple examples.
• we can save the tokenizer along the model directly inside Trainer for the intermediary checkpoints, so it a checkpoint folder can be used directly with our scripts when resuming an interrupted training.

As for the bad part, it's just that it adds a new argument to Trainer.

[codecov]

Codecov Report

Merging #6689 into master will decrease coverage by 1.73%.
The diff coverage is 55.55%.

@@            Coverage Diff             @@
##           master    #6689      +/-   ##
==========================================
- Coverage   78.98%   77.24%   -1.74%     
==========================================
  Files         156      156              
  Lines       28398    28405       +7     
==========================================
- Hits        22429    21941     -488     
- Misses       5969     6464     +495     

Impacted Files	Coverage Δ
src/transformers/trainer.py	53.66% <55.55%> (-0.13%)	⬇️
src/transformers/modeling_tf_albert.py	21.47% <0.00%> (-69.44%)	⬇️
src/transformers/tokenization_xlm.py	16.26% <0.00%> (-66.67%)	⬇️
src/transformers/pipelines.py	25.63% <0.00%> (-54.32%)	⬇️
src/transformers/optimization.py	58.88% <0.00%> (-36.67%)	⬇️
src/transformers/modeling_tf_gpt2.py	65.68% <0.00%> (-29.33%)	⬇️
src/transformers/optimization_tf.py	33.33% <0.00%> (-24.33%)	⬇️
src/transformers/modeling_tf_auto.py	48.79% <0.00%> (-18.08%)	⬇️
src/transformers/data/processors/squad.py	13.76% <0.00%> (-14.38%)	⬇️
src/transformers/modeling_auto.py	64.36% <0.00%> (-14.37%)	⬇️
... and 13 more

---

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 abc0202...54feec2. Read the comment docs.

[LysandreJik]

I like this a lot and feel like it should have been like this from the start! Since we consider models effectively as model-tokenizer pairs, I think it makes a lot of sense for the trainer to handle both.

Especially with regards to saving/reloading, which has always been an issue with users not understanding how to reload from checkpoints as the tokenizers were not saved in the same folder.

[jplu]

Nice! I like it. Ok for me to do the same on the TF one 👍