Implement per-pixel linked list for OIT

[beicause]

Objective

The current OIT stores viewport-sized fragments per layer. It uses much more memory than it can be.

Solution

Implements per-pixel linked list for OIT, which saves memory and can handle more layers. The implementation references https://github.com/KhronosGroup/Vulkan-Samples/tree/main/samples/api/oit_linked_lists

Testing

Tested with the order_independent_transparency example. I also added a new scene in it.

Details

[IceSentry]

This is awesome. Thank you so much for working on this. Sorry it took so long for me to review, I got sick in the same week you opened the PR and haven't had time to come back to it since.

This is very close to what I had in mind as a follow up to my original OIT impl so I'm really happy to see it in action.

I managed to review the PR because I'm very familiar with OIT but to make the diff a bit simpler to follow I would suggest adding the depth prepass support in a separate PR to the current OIT impl. This way the linked list PR will be a bit easier to follow since it won't be mixed with the depth prepass changes.

[IceSentry]

@goodartistscopy btw, we highly encourage community reviews. Since it seems like you already looked over all the changes feel free to leave a review/approve.

[IceSentry]

We need 2 reviews for a maintainer to look at it and merge it.

[tychedelia]

There appears to be a memory leak in the example on macOS, but that's on main so not blocking here. Looks great! Thanks for your work @beicause

[tychedelia]

Note for other reviewers, was curious if contention on this is ever a source of concern, but seems that drivers very well optimize this case and so it's preferred over more complicated optimizations.

[goodartistscopy]

Contention on atomic_counter ("the allocator") is highest at the draw/accumulation phase. In the resolve shader, each thread has it's own head (and list of nodes), so no contention. However all threads also reset atomic_counter non atomically and I wonder if its very pedantically UB.

[beicause]

This references https://github.com/KhronosGroup/Vulkan-Samples/blob/6a4d8b0552df04aad581c746533a03db95ca5012/shaders/oit_linked_lists/combine.frag#L145-L147

[goodartistscopy]

Aside from correctness, I wonder if it costs the full bandwidth too ? Like, would

let screen_index = u32(floor(in.position.x) + floor(in.position.y) * view.viewport.z);
if screen_index == 0 {
    atomic_counter = 0u;
}

be more efficient ?

[beicause]

There appears to be a memory leak in the example on macOS

I think it's due to OitBuffers is never released if OIT is disabled. It should be resolved.

[goodartistscopy]

• atomic_counter is a non informative name (allocator or next_node would be better)
• I'll put the clean up of the shader loops along with #22781 but would be better included here

[IceSentry]

I'll put the clean up of the shader loops

I'd prefer if this was in a self contained small PR. Large PRs always take a long time to merge just because it's harder to review but keeping things small makes the process faster.