[s2s] add BartTranslationDistiller for distilling mBART

[sshleifer]

New class BartTranslationDistiller does the same distillation method as SummarizationDistiller, but computes BLEU scores instead of ROUGE scores. It also accepts --src_lang and --tgt_lang arguments from the command line.

There is one strong checkpoint already posted at sshleifer/distillmbart-12-6/. I will post more in the coming days.

[codecov]

Codecov Report

Merging #6363 into master will decrease coverage by 1.55%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master    #6363      +/-   ##
==========================================
- Coverage   79.93%   78.37%   -1.56%     
==========================================
  Files         153      148       -5     
  Lines       27888    27196     -692     
==========================================
- Hits        22293    21316     -977     
- Misses       5595     5880     +285     

Impacted Files	Coverage Δ
src/transformers/modeling_tf_mobilebert.py	24.55% <0.00%> (-70.09%)	⬇️
src/transformers/tokenization_reformer.py	41.66% <0.00%> (-40.00%)	⬇️
src/transformers/tokenization_marian.py	92.17% <0.00%> (-7.00%)	⬇️
src/transformers/modeling_tf_albert.py	81.74% <0.00%> (-6.61%)	⬇️
src/transformers/modeling_tf_electra.py	94.24% <0.00%> (-1.84%)	⬇️
src/transformers/modeling_longformer.py	89.21% <0.00%> (-1.64%)	⬇️
src/transformers/modeling_tf_bert.py	95.48% <0.00%> (-1.10%)	⬇️
src/transformers/configuration_bart.py	93.75% <0.00%> (-0.49%)	⬇️
src/transformers/modeling_tf_auto.py	66.66% <0.00%> (-0.21%)	⬇️
src/transformers/tokenization_bart.py	95.77% <0.00%> (-0.18%)	⬇️
... and 21 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 be1520d...0718179. Read the comment docs.

[sgugger]

LGTM!