break cyclic meta-layer recursion in loader_add_meta_layer

[aizu-m]

Found while fuzzing the loader settings file with malformed layer manifests.

==ERROR: AddressSanitizer: stack-overflow
    #3 loader_add_meta_layer     loader.c:1493
    #5 loader_add_meta_layer     loader.c:1493
    #6 loader_add_implicit_layer loader.c:1454
    #7 loader_add_meta_layer     loader.c:1493
    ...

Two meta-layers whose component_layers list each other recurse loader_add_meta_layer -> loader_add_implicit_layer -> loader_add_meta_layer and never stop. The stack runs out inside vkCreateInstance.

verify_all_meta_layers already catches these cycles and drops the offending layers, but it only runs over layers found through normal discovery. Layers named in vk_loader_settings.json arrive through get_settings_layers and never reach that pass, so the cycle is still live once expansion starts. b56f871 took the recursion out of loader_find_layer_name_in_meta_layer for the same reason; this sibling was missed.

The guard is a per-layer is_being_expanded flag, set on entry to loader_add_meta_layer and cleared once its components are walked. Re-entering a layer already on the expansion stack logs a warning and returns rather than recursing. A well-formed meta-layer that pulls in another meta-layer still expands exactly as before.

Added SettingsFile.CyclicMetaLayerComponentDoesNotRecurse, which points two settings meta-layers at each other. It overflows the stack under ASan without the guard and passes with it.

[ci-tester-lunarg]

Author aizu-m not on autobuild list. Waiting for curator authorization before starting CI build.

[ci-tester-lunarg]

Author aizu-m not on autobuild list. Waiting for curator authorization before starting CI build.

[charles-lunarg]

OSS-Fuzz found a similar crash that I fixed by simply disabling meta layers from enabling other meta layers.

Since meta-layers aren't necessary any more, I am content with removing the ability for meta layers to enable other meta layers. That feature was never required, only existed as a fallout of 'yeah I guess this could work'. Having variables that live in the layer properties that mutate as the layers are loaded is a counter intuitive design.

In other words, I'm happy to take this PR as it, but would also would be happy with a 'remove this feature' PR. I am happy to make that PR as well, since fixing this bug is already a big help.

[ci-tester-lunarg]

CI Vulkan-Loader build queued with queue ID 38308.

[ci-tester-lunarg]

CI Vulkan-Loader build # 3630 running.

[ci-tester-lunarg]

CI Vulkan-Loader build # 3630 passed.