Support arrays of TLAS bindings

[MarijnS95]

Depends on #1245

Summary

Mirror the existing buffer/texture array pattern (ArraySize plus an ArraySize-driven YAML field) for top-level acceleration structures so shaders can declare RaytracingAccelerationStructure Scenes[N] and bind N distinct TLASes through a single resource entry.

• Schema (Support/Pipeline.{h,cpp}): TLASDesc gains ArraySize (default 1) and reshapes Instances from SmallVector<InstanceDesc> to SmallVector<SmallVector<InstanceDesc>>. MappingTraits<TLASDesc> dispatches on ArraySize exactly like setData() does for CPUBuffer: flat Instances: [...] for the scalar case, list-of-lists for arrays, with an ActualSize != ArraySize validation error on mismatch. Resource::getArraySize() returns TLASPtr->ArraySize for AS resources (moved out-of-line so it can dereference the still-forward-declared TLASDesc).
• createAS(): takes uint32_t InstanceCount directly — pure single-create that sizes and allocates one TLAS. No Pipeline / Resource / InvocationState parameters. The multi-create (createBuffers / createResources) loops TD.ArraySize and pushes one bundle entry per element plus N handles into InvocationState::TLASes, which becomes StringMap<SmallVector<unique_ptr<AccelerationStructure>>> (one vector per TLASDesc::Name, sized to ArraySize).
• buildPipelineAccelerationStructures() walks P.AccelStructs.TLAS and, for each name with a pre-allocated vector, builds one TLASBuildRequest per element using TD.Instances[Elt] and Handles[Elt].
• Vulkan: descriptor write iterates the bundle's ResourceRefs to fill N entries of pAccelerationStructures and sets descriptorCount = R.getArraySize() so the descriptor set sees the full array. Pool sizing (ASDescriptorCount) and binding layout (ResourceBinding.DescriptorCount) already used R.getArraySize() for AS resources — they just start returning the right number once TLASDesc::ArraySize flows through.
• DX12: bindSRV()'s AS branch already loops ResBundle; with the new multi-entry bundle it now writes N RAYTRACING_ACCELERATION_STRUCTURE SRVs into consecutive heap slots automatically. Heap sizing already uses getDescriptorCountWithFlattenedArrays().
• Metal: the AS descriptor-binding loop now builds one IRRaytracingAccelerationStructureGPUHeader + instance-contributions buffer pair per array element and writes one descriptor entry per element via IRDescriptorTableSetAccelerationStructure(). MarkASResident descends into the per-name vector.

New test test/Feature/InlineRT/tlas-array.test declares Scenes[2], each TLAS carries one triangle instance with a distinct InstanceID (10 and 20), and the shader writes each CommittedInstanceID() into Output[i]. Gated on acceleration-structure, XFAIL: Clang.

Test plan

Local on an NVIDIA RTX 3060:

• Linux Vulkan (native offloader)
• Linux D3D12 (Wine + vkd3d-proton + cross-compiled offloader.exe)
• Windows Vulkan (native offloader.exe)
• Windows D3D12 (native offloader.exe)

CI (RT-capable runners):

• windows-nvidia D3D12 (RaytracingTier 1.2)
• windows-intel VK (VK_KHR_ray_tracing_pipeline)
• macOS Metal (supportsRaytracing)

[bogner]

Explaining why this is out of line isn't terribly useful I think. If somebody for some reason really wanted it to be inline they'd figure it out in a hurry.

Suggested change

	uint32_t getArraySize() const; // out-of-line: needs complete TLASDesc.
	uint32_t getArraySize() const;