Improve depth sort

[jatentaki]

Current implementation uses bitonic sort which is less efficient (but simpler to implement) than alternatives like radix sort. Improvements can include both replacing bitonic sort with radix sort and perf improvements to bitonic itself:

• Replacing the array of buffers with a simple buffer and using dynamic offsets
• Using shared memory within the compute shader itself
• Others

[julien-blanchon]

Look like a task for the magic: @dakenf