birefnet: dont dump the full relative position bias for window attention - it's really big after all (800MB for full model) - compute it from bias table and window index like in pytorch