[Opensora-PKU] v1.2.0 training and inference

[wtomin]

Support:

1. run vae reconstruction, t2i and t2v inference;
2. train stage 1 ~ 5;
3. Training with zero2, zero3, and sequence parallelism;
4. Training in pynative mode.
   Overfitting experiment was conducted for stage 3 (29x480p). The overfitting loss curves and generated videos were similar to torch's overfitting experiment.