C2s attention returns genenames and optimisation

[martinaoliver]

Three updates for optimisation and compatibility with attention map analysis

1. get_embeddings function now also returns gene_names_list (order of gene names ranked by tokenizer).
2. _aggregate_token_to_word_attention has GPU implementation for speedup
3. Only process layer selected with emb_layer
   ***NOTE: This function is not being called in bioagents for inference with c2s - is only being used to retrieve attention maps