[SYCL][Bindless] Fix allocation for exportable Vulkan interop images

[mateuszpn]

depth_format failed on Windows Intel GPUs not because of a SYCL/UR bug but because the test built an invalid/mis-described VkImage. Three defects, all things the passing sibling tests (vulkan_setup.hpp) already do correctly:

• Under-sized allocation — memory was sized as w*h*sizeof(float) instead of the driver's VkMemoryRequirements::size. That's smaller than required on any GPU that pads tiled images (hard VUID violation); hidden on BMG (zero padding), surfaced on DG2 (65536 < 73856).

• No dedicated allocation — used only when the driver reported it required, but BMG under-reports for external memory → pool-allocated → exported handle describes the wrong layout → SYCL reads land at wrong offsets.

• Missing STORAGE_BIT — image was transfer-only, but the kernel accesses it as a storage image.

Fix: in the shared vulkan_common.hpp, use the real memory-requirement size and use dedicated allocation for any exportable image; in the test, add STORAGE_BIT and thread the real size into the SYCL import descriptor. mipmaps.cpp (the only other consumer) is strictly more correct too.

Result: depth_format passes 3/3 on BMG (Arc B580, 32.0.101.8425) → BMG XFAIL (#21986) removed. DG2 XFAIL (#21985) kept — this is a driver bug, not a test issue.

Details and per-line rationale in the inline comments

[dyniols]

@mateuszpn Please, ping me when PR will be ready. I will review it for intel/bindless-images-reviewers group.

[dyniols]

@mateuszpn Could you create issue on driver team if you believe that failure on DG2 is because of driver bug? If not let me know I can do it. Since I would try get resolution on #21985 and only test that prevents it vulkan_interop/depth_format.cpp‎.