Remove outdated BERT tips

[JetRunner]

Why remove the tips:

• BERT is a model with absolute position embeddings so it's usually advised to pad the inputs on
  the right rather than the left.

Yes but since we don't provide an option to pad from the left I think it's not necessary.

• BERT was trained with a masked language modeling (MLM) objective. It is therefore efficient at predicting masked
  tokens and at NLU in general, but is not optimal for text generation. Models trained with a causal language
  modeling (CLM) objective are better in that regard.

No. T5 & BART proved it wrong.

• Alongside MLM, BERT was trained using a next sentence prediction (NSP) objective using the [CLS] token as a sequence
  approximate. The user may use this token (the first token in a sequence built with special tokens) to get a sequence
  prediction rather than a token prediction. However, averaging over the sequence may yield better results than using
  the [CLS] token.

No. [CLS] can do learnable self-attention pooling, which is way much better than parameter-free average pooling especially when fine-tuned. (w.r.t. SentenceBERT)

[codecov]

Codecov Report

Merging #6217 into master will decrease coverage by 0.64%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master    #6217      +/-   ##
==========================================
- Coverage   79.53%   78.88%   -0.65%     
==========================================
  Files         146      146              
  Lines       26586    26586              
==========================================
- Hits        21145    20973     -172     
- Misses       5441     5613     +172     

Impacted Files	Coverage Δ
src/transformers/modeling_outputs.py	100.00% <ø> (ø)
src/transformers/tokenization_xlm.py	16.26% <0.00%> (-66.67%)	⬇️
src/transformers/tokenization_marian.py	68.14% <0.00%> (-25.67%)	⬇️
src/transformers/generation_tf_utils.py	86.71% <0.00%> (+0.25%)	⬆️
src/transformers/file_utils.py	80.30% <0.00%> (+0.25%)	⬆️
src/transformers/tokenization_dpr.py	57.65% <0.00%> (+4.50%)	⬆️
src/transformers/tokenization_ctrl.py	96.11% <0.00%> (+17.47%)	⬆️

---

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 b6b2f22...8a7c591. Read the comment docs.

[sgugger]

I'd personally leave the first two tricks (users may have their custom script for padding and BERT is not good at language generation). For the third, as discussed, I agree.
The docstrings of BaseModelOutputWithPooling should be changed accordingly, but I can take care of that in a separate PR.

[JetRunner]

@sgugger I've restored tips no.1 and updated tips no.2. Also took care of BaseModelOutputWithPooling.

[sgugger]

Thanks a lot! This is good to go IMO.