Release optimizer memory and fix legacy tokenization

[helpmefindaname]

This PR does 2 little things:

• it makes it easier to clean up the gpu memory, by cleaning gradients and deleting the reference of the optimzier. Hence torch.cuda.empty_cache() will be able to clean up more parameters.
• some (slow) tokenizers like RobertaTokenizer seem to add a [SEP] token at the end, breaking the legacy subword token mapping, which is fixed by allowing that specific token

[helpmefindaname]

With those changes, I managed to train XLM-Roberta-Large (>2GB model) with adam and full precision (4x the memory requirement) on my 6GB laptop graphic card, using transformer-smaller-training-vocab

[alanakbik]

@helpmefindaname thanks for improving this!