Fix Trainer in DataParallel setting

[sgugger]

The new output types seem to break data parallel FYI, see comment on #5671. This is is because of the line

return type(out)(map(gather_map, zip(*outputs)))

in scatter_gather which tries to reconstruct an output of the same type as ours (and fails since it does not provide the necessary arguments). There is no way to fix our ModelOutput to work with this AFAICT.

However, we have the return_tuple argument to fix the issue :-)

[codecov]

Codecov Report

Merging #5685 into master will decrease coverage by 0.20%.
The diff coverage is 25.00%.

@@            Coverage Diff             @@
##           master    #5685      +/-   ##
==========================================
- Coverage   78.11%   77.91%   -0.21%     
==========================================
  Files         146      146              
  Lines       25983    25987       +4     
==========================================
- Hits        20297    20247      -50     
- Misses       5686     5740      +54     

Impacted Files	Coverage Δ
src/transformers/trainer.py	37.84% <25.00%> (-0.12%)	⬇️
src/transformers/modeling_tf_t5.py	44.56% <0.00%> (-46.35%)	⬇️
src/transformers/modeling_tf_gpt2.py	63.55% <0.00%> (-31.78%)	⬇️
src/transformers/generation_tf_utils.py	79.94% <0.00%> (-6.02%)	⬇️
src/transformers/modeling_tf_utils.py	86.92% <0.00%> (-1.97%)	⬇️
src/transformers/modeling_openai.py	82.31% <0.00%> (+1.28%)	⬆️
src/transformers/modeling_tf_roberta.py	93.36% <0.00%> (+49.37%)	⬆️
src/transformers/modeling_tf_openai.py	95.18% <0.00%> (+74.91%)	⬆️

---

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 7fad617...05ec8f6. Read the comment docs.

[thomwolf]

LGTM