Fix: SISR residual block architecture and optimize forward pass

[kamilansri]

Summary

This PR fixes a critical logic error in the SISR model's residual layers and optimizes memory usage during the forward pass.

Changes

• Critical Fix: Removed the final ReLU activation in make_residual_layer.

  • Reasoning: Residual blocks must compute x + Residual(x). If the residual path ends in a ReLU, it can only add positive values to the signal, severely limiting the network's capacity to learn.

• Optimization: Removed x.clone() calls in the forward method.

  • Reasoning: PyTorch layers generate new tensors; explicitly cloning input tensors before passing them to layers is redundant and increases VRAM usage unnecessarily.

• Refactoring: Updated super(SISR, self).__init__() to the modern Python 3 super().__init__() syntax.