Improve UNet Shallow end-to-end performance

[esmalTT]

Summary

The current single-chip end-to-end performance of UNet is 243 fps. In order to improve the end-to-end performance of UNet, we need to do the following things:

1. Reduce input volume: Currently input channels are padded from 4 to 32 because convolution does not support in_channels=4.

   • Support for Conv2D with dims < 32 and dims % 32 != 0 #11955
   • TTNN pad does not support last dimension #12896 - possible workaround

2. Reduce output volume: UNet's output has only a single channel, but because the output tensor TILE layout, we copy an extra 31 channels of data back to host.

   • TTNN to_layout does not support UNet Shallow output shape #12705

3. Improve host read-back performance: based on what we are seeing in the tracy profile, UNet seems to be spending a long time in the host read back

   • UNet Shallow host read-back performance is slow #12837