Add Adapters to DistilBERT

[calpt]

Changes:

• Adds adapter support to DistilBERT models (via mixins in adapter_distilbert.py)
• Adds a flex-head model for DistilBERT (DistilBertModelWithHeads)
• Moved invertible adapters to separate InvertibleAdaptersMixin to improve modularity
• Adjustments in BERT adapters implementation to allow partial reuse for DistilBERT

TODO:

• Wait for merge of (at least) Minor fixes to Fusion code, examples, tests & docs #61
• Add to documentation
• Train adapters on GLUE to validate implementation

Closes #58.

---

Results on GLUE:

Task	Pfeiffer (dev)	Houlsby (dev)
RTE (acc)	61.97 *	60.17 *
MRPC (acc)	85.29 *	85.94 *
STS-B (spearmanr)	85.75 *	85.83 *
CoLA (acc)	51.44 *	51.86 *
SST-2 (acc)	90.36 *	90.71 *
QNLI (acc)	88.5 *	88.67 *
QQP (acc)	89.28	90
MNLI (acc)	81.33	82.08

(*) mean of 3 runs

[hSterz]

Looks good I just have question about the code conventions