Add missing token classification for XLM

[sakares]

The current modeling_xlm.py did not have alike ForTokenClassification class like others, which helps for NER task comparison across all existing models.

now XLMForTokenClassification can be called via:

from transformers import XLMForTokenClassification

model = XLMForTokenClassification.from_pretrained('xlm-mlm-100-1280')

[sshleifer]

LGTM @LysandreJik

[LysandreJik]

Cool, thanks @sakares! I rebased on your branch and added a commit for tests that should be slow (question answering and token classification in auto models)