[Tests] fix attention masks in Tests

[patrickvonplaten]

This PR should fix the flaky test failures of test_modeling_output_equivalence and test_feed_forward_chunking.

I added a new random attention_mask generation function that makes sure that at least one token is attended to.

[codecov]

Codecov Report

Merging #6621 into master will increase coverage by 0.85%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master    #6621      +/-   ##
==========================================
+ Coverage   79.16%   80.02%   +0.85%     
==========================================
  Files         156      156              
  Lines       28217    28217              
==========================================
+ Hits        22339    22581     +242     
+ Misses       5878     5636     -242     

Impacted Files	Coverage Δ
src/transformers/modeling_tf_xlm.py	18.94% <0.00%> (-74.32%)	⬇️
src/transformers/modeling_tf_flaubert.py	24.53% <0.00%> (-63.81%)	⬇️
src/transformers/modeling_roberta.py	77.37% <0.00%> (-19.71%)	⬇️
src/transformers/file_utils.py	82.18% <0.00%> (-0.26%)	⬇️
src/transformers/modeling_utils.py	88.05% <0.00%> (+0.55%)	⬆️
src/transformers/modeling_tf_utils.py	87.29% <0.00%> (+2.60%)	⬆️
src/transformers/generation_tf_utils.py	86.46% <0.00%> (+2.75%)	⬆️
src/transformers/configuration_t5.py	96.42% <0.00%> (+10.71%)	⬆️
src/transformers/modeling_t5.py	83.83% <0.00%> (+12.21%)	⬆️
src/transformers/modeling_tf_t5.py	90.93% <0.00%> (+64.09%)	⬆️
... and 1 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 573bdb0...d61cbf8. Read the comment docs.

[sgugger]

Thanks for digging into this and fixing those flaky tests!