Add first version of the efficient decoder

[Flova]

Proposed changes

• Reorganize the order of upscaling and conv layers so the conv layers operate on a smaller image and do equivalent work
• Look at layer wise profiling results from the TVM profiler on our vulkan deployment
• Reduce number of channels in the decoder and add a 1x1 conv which reduces the channels after the concatenation of the skip connections.
• Optimize the TVM compilation
• Use mixed precision

These changes bring the performance on the AMD Vega ATI 05:00.0 Lucienne (GPU of the Ryzen 7 5700U) with the Vulcan API from ~20 FPS to ~90 FPS. The accuracy in full precision is identical. Half needs to be evaluated.

[jaagut]

What's the status of this? Is it ready for a merger?

[Flova]

I bumped some versions and the ci fails