Allow llama_transformer take embedding as optional argument (#4257)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/4257

Llava run token embedding separately and combine that with
image embeddings as input to text transformer. This PR adds support for
that.

Reviewed By: tarun292

Differential Revision: D59759977

fbshipit-source-id: 45fe8af0bdfc090cb135d3d3f0b8ff3a18fa1b3c