Is your feature request related to a problem? Please describe.
Right now our UNet implementation only supports the standard DoubleConv blocks and plain upsampling. This makes it harder to experiment with common U-Net variants (Residual U-Net, Attention U-Net) that are widely used in medical image segmentation.

Describe the solution you'd like
I’d like to add support for the new components we implemented in components.py:

• UNetResidualConv (residual double conv block)
• AttentionGate + UNetAttentionUpsample (skip connection gating)

The UNet class could accept arguments such as block= and upsample= so users can flexibly switch between the default modules and these new ones.

Describe alternatives you've considered

• Keeping a completely separate ResUNet or AttentionUNet implementation, but that would duplicate most of the code.
• Hard-coding these options into UNet, but that would reduce flexibility.

Additional context
These modules are standard in literature, and integrating them would make mipcandy-bundles much more “out-of-the-box” for segmentation baselines. We already have working implementations, we just need to hook them into the UNet API and add minimal tests.