Enable multi-gpu training when "torch" is chosen as the RETURNN backend

[Judyxujj]

Using the current mpirun to launch the torch distributed data parallel (DDP) training gives error ValueError: Error initializing torch.distributed using env:// rendezvous: environment variable RANK expected, but not set. Therefore in order to enable the multi-gpu training for torch backend, once backend = "torch" is detected in returnn config, ReturnnTrainingJob will now use torchrun to launch DDP training.