OpenGL out, Vulkan in: four decisions

Rung 0 of the Vulkan backend, pull request #2117, brought a second device implementation up on this machine and put the settled decisions in code. volk loads every call through the device's own pointers after creation, so nothing links the loader. The device is created with Vulkan 1.3's dynamic rendering and synchronization2, and a device below 1.3 is refused rather than worked around, since a compatibility path is the stack of layers the fourth constraint rules out. Memory is the backend's own block suballocator rather than VMA: 64 MB blocks per memory type with a free list, a dedicated allocation for anything larger, host-visible blocks mapped once. The one shader source arrived with #2088.