setMightHaveLatentChildren(false) even if a tile has real children.

[kring]

Fixes CesiumGS/cesium-unreal#1529

In #783, I renamed shouldContinueUpdating to mightHaveLatentChildren in an attempt to be more descriptive about what it actually does. I also implemented what I thought was a very slight tweak to the logic for creating explicit children from latent ones.

Previously, when we called updateTileContent on a tile that mightHaveLatentChildren (which is all tiles to start), we would give the TilesetContentLoader associated with the tile a chance to create explicit children. While looking at this in the above PR, I noted this was a bit dangerous in the case that the tile already has explicit children. It could end up clobbering the existing children, which would be very likely to lead to a hard to debug crash.

Of course, none of our current loaders actually have this problem. Our current implicit loaders create one tile at a time, so every tile has no children to start, and so this situation never arises. Still, TilesetContentLoader is meant to be a generic interface, so I thought it was a good idea to defensively avoid this possibility.

So I added a check: if the tile already has real children, don't call the loader to create more. It seemed like a good idea, and I thought it was safe.

Unfortunately this mightHaveLatentChildren flag has a side effect that I didn't consider, and my logic change meant that this flag stayed true for tiles with real children where previously it flipped to false the first time the tiles were accessed. And as a result, some logic in updateDoneState that turned placeholder raster overlays into real ones never executed. And that, in turn, would cause tilesets with raster overlays to never load at all in some cases, as in CesiumGS/cesium-unreal#1529.

The fix in this PR simply makes createLatentChildrenIfNecessary set mightHaveLatentChildren to false when it sees a tile that already has real children. That allows the placeholder replacement logic to run, which unblocks loading.

[kring]

I'm merging this because it fixes a regression that would otherwise be in today's release.