Refactor: Removed CachedVisibilityMap to use previous frame data, fixed another crash.

[Jaskowicz1]

This PR removes the CachedVisibilityMap to use the previous frame's data to check if objects should still be occluded or not.

This saves, at a minimum, 188 bytes of memory. In a scene with 1000 objects (this is quite easy to hit, reminder, each Static Mesh inside of a blueprint is another object), this is roughly ((9*1000)+188)=9188 bytes (9.1KBs) of memory (9 is the size of FPrimitiveSceneID + bool), actual saved memory might be smaller or bigger, depends on usage.

This PR also stops sending every object in the FrameResult's VisibilityMap to the GameThread if it hasn't changed. This improves the frame time quite a lot in a heavy populated world (I've seen about a ~0.2-0.4ms improvement in my test scene), as we no longer send a task for every object in the world, only ones that need to be occluded (and if their visibility doesn't match what the occluder wants).

Closes #1

[Jaskowicz1]

Draft PR whilst I work out why the crash exists.

[Jaskowicz1]

Figured out the cause of the Crash in Packaged Builds.

IndexBuffer.AccessStream16() returns nullptr as any mesh passed through to FOcclusionMeshData doesn't have bAllowCPUAccess set to true.

This defaults to false on all Static Meshes. Meaning cooked builds cannot access data from any IndexBuffer on a static mesh.

All meshes would have to do this:

How wonderful...

[Jaskowicz1]

Whilst people could in theory make all their static meshes change to support this and whatnot, it'll increase RAM but decrease VRAM.

This may be fine tbh (especially for platforms that have shared RAM because you're not actually losing anything, it's just now in a different place).