llama.cpp issue #27734: ~78% Vulkan decode cliff at 131072 context from suballocation fragmentation

RX 7900 XTX / Win11 / build 10603: 40.4 tok/s at ctx 65536 and 40.8 at 98304, then 8.9 tok/s at ctx 131072; GGML_VK_SUBALLOCATION_BLOCK_SIZE=4294967296 restores 40.3 tok/s. GGML_VK_MEMORY_LOGGER=1 shows 21.55 GiB fully in VRAM split into 30+ ~1018 MiB VkBuffers. A second reporter reproduced the fragmentation on an RTX 5070 Ti; a maintainer-side comment (tczank) reproduced the fragmentation on Linux/RADV but NOT the throughput cliff, so the causal claim is contested.