[s2s] wmt download script use less ram

[stas00]

a few enhancement to #6403

the main change:

• rewrite not to load 100GB into RAM - wmt19 is huge!

and then a few small things:

• replaced the default dataset to wmt16 as it's much much smaller than wmt19 to experiment with (also it seems that at least wmt19-ru-en is missing a test dataset, while wmt16-run-en has it)
• added lang defaults so it's easy to start experimenting with
• moved tqdm to where a detailed progress can be seen
• added some extra notes

@sshleifer

[codecov]

Codecov Report

Merging #6405 into master will decrease coverage by 2.10%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master    #6405      +/-   ##
==========================================
- Coverage   80.10%   78.00%   -2.11%     
==========================================
  Files         149      149              
  Lines       27680    27680              
==========================================
- Hits        22173    21591     -582     
- Misses       5507     6089     +582     

Impacted Files	Coverage Δ
src/transformers/optimization.py	28.94% <0.00%> (-67.11%)	⬇️
src/transformers/pipelines.py	26.98% <0.00%> (-52.81%)	⬇️
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/modeling_auto.py	63.95% <0.00%> (-14.54%)	⬇️
src/transformers/data/processors/squad.py	13.76% <0.00%> (-14.38%)	⬇️
src/transformers/modeling_tf_gpt2.py	65.68% <0.00%> (-6.16%)	⬇️
src/transformers/modelcard.py	82.71% <0.00%> (-2.47%)	⬇️
src/transformers/modeling_distilbert.py	96.19% <0.00%> (-1.64%)	⬇️
src/transformers/configuration_utils.py	95.20% <0.00%> (-1.37%)	⬇️
... and 10 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 b9ecd92...de72b0a. Read the comment docs.

[stas00]

further testing showed that chunking doesn't make much of a difference - writing one record at a time is almost as fast as writing in chunks of 10K records - I think it's the reading that's the bottleneck here, which we can't optimize. So I removed the chunking from this PR.

wmt19-ru-en with 37M records converted in ~40mins on my machine using this PR.