replace `torch.arange(...)` with lookup

[YYYasin19]

Currently the pos. embeddings are always re-calculated with torch.arange which can easily be replaced by an initial tensor that is then indexed everytime