Fix __init__ to call original CrossEncoder initialization

Remove lazy loading and load model on instantiation

Remove logging statements

Use torch.inference_mode() and disable gradient checkpointing

Update README.md

releasing zerank-2