UPSTREAM PR #17624: vulkan: set all memory allocations to high priority

[loci-dev]

Mirrored from ggml-org/llama.cpp#17624

For #17605, though I'm not sure whether it'll help.

[loci-review]

Explore the complete analysis inside the Version Insights

Performance Review Summary

PR #373: Vulkan Memory Priority Implementation

This PR adds Vulkan memory priority support by enabling VK_EXT_memory_priority extension and setting all memory allocations to maximum priority (1.0f). The changes modify device initialization and buffer allocation logic in ggml-vulkan.cpp.

Performance Impact: No measurable performance changes detected across all binaries. All metrics show 0.0% change in response time and throughput. The modifications are runtime configuration changes that do not alter computational paths or execution logic.

Power Consumption: No change detected. All binaries maintain baseline power consumption (libggml-cpu.so: 115,347 nJ).

Inference Impact: No impact on tokens per second. Core inference functions (llama_decode, llama_encode, llama_tokenize) are unaffected by these Vulkan backend initialization changes.