Bindless Descriptor fixes and optimizations

[Firestar99]

• added DescriptorSet::invalidate() to make vulkano forget about resources that bound to a descriptor_set, so they can be freed. As (normal) descriptors are validated when bound, this is entirely safe to do.
• disable validation for individual descriptors if the descriptor set is UPDATE_AFTER_BIND or PARTIALLY_BOUND, making them actually usable
  • I'm not sure these changes are entirely safe. For it to be, we'd need to validate the descriptor set is properly populated just before flushing any commands to the GPU. And I'm not sure if that's done already, or where it should be added. We could ofc also decide to just leave these two flags as unsafe, somehow.
• some small Descriptor optimizations:
  • More SmallVec use in DescriptorPool to improve variable descriptor count allocation. It only ever allocates one descriptor set at a time, i.e. allocate_infos only ever has one entry in that code path. Annoyed me and just had to fix :D
  • Descriptor updates with no writes or copies return early instead of calling into vulkan. Particularly useful when creating a new descriptor, writing nothing, and updating it properly later.

1. Update documentation to reflect any user-facing changes - in this repository.

2. Make sure that the changes are covered by unit-tests.

3. Run cargo clippy on the changes.

4. Run cargo +nightly fmt on the changes.

5. Please put changelog entries in the description of this Pull Request
   if knowledge of this change could be valuable to users. No need to put the
   entries to the changelog directly, they will be transferred to the changelog
   file by maintainers right after the Pull Request merge.

   Please remove any items from the template below that are not applicable.

6. Describe in common words what is the purpose of this change, related
   Github Issues, and highlight important implementation aspects.

Changelog:

### Additions
* added `DescriptorSet::invalidate()` to make vulkano forget about resources that bound to a descriptor_set, so they can be freed

### Bugs fixed
* fixed descriptor sets with `UPDATE_AFTER_BIND` or `PARTIALLY_BOUND` being wrongly validated on bind

[Firestar99]

I've turned this PR into a draft, as I've added DescriptorSet::invalidate() to it and haven't gotten around to extensively test it tested it and it works :D

[marc0246]

I think you already know this, but to anyone reading this, I would like to murder DescriptorSet before the next release. There's no point having all these collections in so many places that just keep references to Vulkan objects, cloning and dropping Arcs and allocating and deallocating the Vecs / HashMaps / whatever for them each and every frame multiple times, when there is one central resources collection that tracks everything in one place. I'm willing to merge this regardless if for nothing else but to not inconvenience you, but I wouldn't start designing something new around this invalidation API.

[marc0246]

Regarding PARTIALLY_BOUND, I think we decided that all shader executions are unsafe, and it is documented that the shader must not access descriptors that haven't been bound. However, the logic here is not correct. It's only PARTIALLY_BOUND that permits this, and not UPDATE_AFTER_BIND.

[Rua]

If UPDATE_AFTER_BIND is enabled, then it's possible that the resources in the descriptor set, at the time the command is recorded, are not the same as those that will be bound when the command buffer is actually executed. So any record-time checks on the bound resources are probably superfluous?

[Firestar99]

So any record-time checks on the bound resources are probably superfluous?

Yes they are, so for now I've just disabled them. Which means they are indeed not validated at all, and honestly I'd keep it that way as I don't think modding that into the old sync has any benefit. Rather invest that time into the new sync. We may want to make these flags unsafe, but wouldn't know how.