need better token_type_id processing on transformer GPT2Model

### Feature request in the GPT2Model code, I see that the token_type_id value is embeded with the same **wte** of input tokens. so that the two token type id values: 0, 1, are in collision with input sentence tokens. We need a more nature design here so that we do not have to reproduce vocab_size and sentence token values when trained with the `token_type_ids` inputs. ### Motivation make it simply the same vocab size and sentence token inputs for training GPT2Model with or without `token_type_ids` input. ### Your contribution maybe another embedding layer can be used for the token_type_id embeddings.