Opensora-PKU v1.3.0 : merge to v0.3.0-dev

[wtomin]

The purpose of this PR is to merge the Opensora-PKU v1.3.0 code into the branch v0.3.0-dev.

I directly copied the opensora_pku/ folder from the master branch to the corresponding folder of the v0.3.0-dev branch. Three additional files required for test are listed as: prompt_list_opendv_16, train_t2v_stage3_finetune_opendv_16.sh, sample_opendv_16.sh.

I have run all steps in the testing procedures with this PR. All scripts have passed. I made a minor modification to train_t2v_diffusers.py:

        params_func = inspect.signature(prepare_train_network).parameters
        if "optimizer_parallel_group" in params_func:
            kwargs = {"optimizer_parallel_group": GlobalComm.WORLD_COMM_GROUP}
        elif "op_group" in params_func:
            kwargs = {"op_group": GlobalComm.WORLD_COMM_GROUP}
        else:
            kwargs = {}
        net_with_grads = prepare_train_network(
            latent_diffusion_with_loss,
            optimizer,
            zero_stage=args.zero_stage,
            comm_fusion=comm_fusion_dict,
            scale_sense=loss_scaler,
            drop_overflow_update=args.drop_overflow_update,
            gradient_accumulation_steps=args.gradient_accumulation_steps,
            clip_grad=args.clip_grad,
            clip_norm=args.max_grad_norm,
            ema=ema,
            **kwargs,
        )

This is because pretrain_train_network from mindone/trainers/zero.py has updated its key arguments names: in master branch, it's optimizer_parallel_group, but in 0.3.0-dev branch, it's op_group.