- replace per-expert mask scan with sort+bincount grouped dispatch - carry router stats in forward output instead of module state - keep MoE diagnostics working under DDP/FSDP wrappers - remove unused _load_balancing_loss helper
- replace per-expert mask scan with sort+bincount grouped dispatch - carry router stats in forward output instead of module state - keep MoE diagnostics working under DDP/FSDP wrappers - remove unused _load_balancing_loss helper