update with V2 models

update with 12L and 20L i4096 gc95M models, multitask and quantiz code

Adding `safetensors` variant of this model (#209)