Fix #3305: run_ner only possible on ModelForTokenClassification models

[julien-c]

Also, #3305 breaks (if i'm not mistaken) the ability to run the example script from a pip-installed instance of transformers (vs. from an instance installed from source) (This PR does not fix that second issue)

[codecov-io]

Codecov Report

Merging #3335 into master will decrease coverage by 0.06%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master    #3335      +/-   ##
==========================================
- Coverage   77.63%   77.57%   -0.07%     
==========================================
  Files         100      100              
  Lines       16943    16943              
==========================================
- Hits        13154    13143      -11     
- Misses       3789     3800      +11

Impacted Files	Coverage Δ
src/transformers/modeling_tf_utils.py	86.4% <0%> (-1.97%)	⬇️

---

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 20139b7...7fc00b6. Read the comment docs.

[julien-c]

Ok I'll merge this @srush @LysandreJik as it's more correct (feel free to let me know of your feedback anyways)

[srush]

Thanks. Why does it break the pip version?

[julien-c]

Because MODEL_MAPPING from modeling_auto is not exposed in the package's __init__.py