Revise Train Scripts [Latte and DiT]: avoid setting positional embedding to trainable parameters

[wtomin]

For DiT and Latte models, the positional embedding tensor is initialized as a Parameter that does not require gradient updates.

However, in the training script, I accidently set all parameters to trainable. It is not desirable for positional embeddings, because they are initialized and then fixed during training.

[zhtmike]

maybe just change to tensor and skip load the pos embed parameter, prevent accident change all parameter to be trainable

[wtomin]

maybe just change to tensor and skip load the pos embed parameter, prevent accident change all parameter to be trainable

Now the pos_embed and temp_embed (in latte only) are Tensors instead of Parameters.