llama.cpp b9969: Vulkan fix stops llama-cli crashes on Adreno GPUs

llama.cpp b9969 fixes a Vulkan crash on Adreno GPUs that broke llama-cli on longer prompts with q4_0 quantized models — one of the most common formats for local LLM inference on Android.

What changed:

• Vulkan matmul routing — large matmuls now route to medium tile on Adreno (#24877)
• Prompt crash fix — llama-cli no longer breaks on longer prompts with q4_0 networks
• Root cause — insufficient shared memory on Adreno for the previous tile size
• Adreno cleanup — removed unused Adreno device from the pipeline
• Small pipeline — matmul updated for the small pipeline path

The crash hit real workloads, not edge cases. q4_0 is a default quantization pick for on-device inference, and longer prompts are the norm for chat and agent loops — exactly where Adreno GPUs live (Qualcomm Snapdragon phones and tablets).

Build coverage spans macOS (arm64, x64), Linux (CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL FP32/FP16), Windows (CPU, CUDA 12/13, Vulkan, OpenVINO, SYCL, HIP, OpenCL Adreno), Android (arm64 CPU), and openEuler (310p, 910b ACL Graph). General openEuler is disabled.

Available now — prebuilt binaries for all active targets are on the b9969 release page.