[bugfix] LoRA + TP>1 + sequence_parallel: sum grads of the replicated LoRA factor over the TP group (exported adapter had last-layer linear_proj.lora_B == 0) - #173