CUDA kernels incompatible with standard PyTorch device movement with 4bit/8bit, necessitating device-specific handling (#416)

precommit formatting

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

move dicts to init

Add function to get number of model embeddings (#364)

fix cell state gene embeddings bug (#345)

add load model for train and fix validate anchor gene error

update dependencies, reinstate compatibility with python<3.9 with typing for List

Add functions for extracting gene embeddings, move state_embs_dict outside isp, fix bugs in isp