Tighten up race condition risk in VulkanManager.

Checking for mDevice != null risks using inconsistent state if the
HardwareBufferUpload thread and the render thread race in setting up
vulkan. Instead, use an atomic bool and std::call_once to manage
initiaizing VulkanManager instances.

Bug: 280178674
Test: builds, boots
Change-Id: Ic0a1c3ae1939ece536eb57de369232b213236d11