Vulkan OOM logic can prevent creating large buffers #9206
Description
Test case for this is webgpu:api,validation,buffer,create:limit:* on Radeon Pro W7600.
It appears that because there are two heaps and candidate allocations are checked against both of them, allocations larger than the host-visible heap will always be rejected.
wgpu/wgpu-hal/src/vulkan/device.rs
Line 850 in 9ed7737
(It is also possible to exhaust the host-visible heap by running the CTS test webgpu:api,operation,adapter,requestAdapter:*, but in that case I think it could really be exhausted by rapid creation of many devices.)
The topic of large buffers is also discussed in #8105.