DiffusionModelEncoder output layer is set outside of the initiation and left out of trainable parameters and set to the wrong device

[virginiafdez]

Describe the bug
This issue has been observed in the tutorials: Project-MONAI/tutorials#2030 (comment).

Expected behavior
The out layer should not be set in the forward loop because:

• The device is wrong in this scenario
• The parameters from this layer are not included in the optimizer trainable args.

Additional context
This was caused by fix #8514.