[Feature]: Adding an encoder-decoder model where encoder and decoder are both transformers

[chelseagzr]

Problem statement

Encoder-decoder models, where both the encoder and decoder are transformers, have proven to be highly effective for seq2seq tasks, such as translation. It would be extremely helpful if Flair supports this type of model.

Solution

It would be nice if a user can easily create an encoder-decoder model by specifying the name of the encoder (e.g., xlm-roberta-base) and the name of the decoder (e.g., GPT2). Developers can make use of the EncoderDecoderModel class from transformers module to facilitate this.

Additional Context

No response