Refactored the memory allocation from CAssetConverter::reserve so that it can be used during the conversion as well (needed for Acceleration Structure compactions)

Finalize the TLAS compaction

Also:
- made the allocator used during `reserve` overridable (no longer are using plain device)
- remembered that Overflow Submit can break a commandbuffer, so if we want to use debug markers, we need to use them differently.

Author: devsh

examples_tests

@@ -891,6 +891,8 @@ class CAssetConverter : public core::IReferenceCounted
 			asset::IShaderCompiler::CCache* readShaderCache = nullptr;
 			asset::IShaderCompiler::CCache* writeShaderCache = nullptr;
 			IGPUPipelineCache* pipelineCache = nullptr;
+			// optional, defaults to the device
+			IDeviceMemoryAllocator* allocator = nullptr;
         };
 		// Split off from inputs because only assets that build on IPreHashed need uploading
 		struct SConvertParams