WAN-video: support training with batch size > 1

[hnyu]

Currently, if we want to set batch size > 1 for finetuning WAN, there will be dim0 size mismatch at

x = torch.cat([x, y], dim=1)  # (b, c_x + c_y, f, h, w)

in wan_video_dit.py. The reason is that the code is hardcoded to assume batch_size=1 when processing "context", "clip_feature" and "y" in training_step() of train_wan_t2v.py. We should actually use squeeze(1).