Vulkan: Fix the memory allocation change in #17110

[0cc4m]

I forgot to exit the second loop early in #17110. This should fix #17117.

[Goldenkoron]

On strix halo, trying to load any model on windows with 96gb set in UMA will give a memory buffer error, seemingly no matter the allocation size.

With 64gb set, models will try to load up to 96gb but if the memory actually exceeds 64gb it will give a device out of memory error (even if there is room in Windows shared memory, it never uses it).