Add a transformers PreTrainedModel for MSA Pairformer

This allows for downstream use of MSA Pairformer with HuggingFace libraries like PEFT and TRL.

msa_pairformer_tokenizer = AutoTokenizer.from_pretrained('yoakiyama/MSA-Pairformer')
msa_pairformer = MsaPairformer.from_pretrained('yoakiyama/MSA-Pairformer')

inputs = msa_pairformer_tokenizer(msa)
results = msa_pairformer(inputs['input_ids'], inputs['attention_mask'])

Github PR:
https://github.com/yoakiyama/MSA_Pairformer/pull/4