Fold batches into channels and use grouped convolutions in UNet Shallow

[esmalTT]

Summary

• Improve UNet performance by folding batches into channels and use grouped convolutions in UNet Shallow.
• Residual concatenation handled with ttnn.concat by passing groups parameter (available since 4fc773a).
• This change depends on: Add support for multi-device tensors in grouped convolution weight preprocessing #14914.

Checklist

• Post commit CI passes
• Model regression CI testing passes (if applicable)
• Device performance regression CI testing passes (if applicable)
• New/Existing tests provide coverage for changes

[esmalTT]

Waiting for #14914 to be merged before removing draft status