Speedup torch summarization tests

[sshleifer]

Speedup torch summarization tests by using small models that are faster to download and instantiate.

[sshleifer]

Non slow test speed before change:

[patrickvonplaten]

There is a tiny TF T5 model now as well via:
model = TFAutoModelWithLMHead.from_pretrained("patrickvonplaten/t5-tiny-random")

[codecov-io]

Codecov Report

Merging #3663 into master will decrease coverage by 0.01%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master    #3663      +/-   ##
==========================================
- Coverage   78.03%   78.02%   -0.02%     
==========================================
  Files         104      104              
  Lines       17708    17708              
==========================================
- Hits        13819    13817       -2     
- Misses       3889     3891       +2     

Impacted Files	Coverage Δ
src/transformers/modeling_tf_utils.py	92.96% <0.00%> (-0.17%)	⬇️
src/transformers/modeling_utils.py	91.97% <0.00%> (-0.13%)	⬇️

---

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 0a9d09b...4751188. Read the comment docs.

[julien-c]

LGTM, but FYI for context, I think the reason we were using the real models was that we intended to do integration testing: in _test_mono_column_pipeline we only test equality of keys but we could have tested equality (or closeness) of values.

[sshleifer]

Makes sense @julien-c . I'd be happy to add some @slow integration tests and try to fulfill the original intent