Wiring Android's Vulkan Compute to a Quantized On-Device Diffusion Model for Real-Time Image Segmentation

Deep dive into using Vulkan compute shaders directly (bypassing NNAPI) for a quantized segmentation model — covering descriptor set management, push constants for dynamic tensor dims, ping-pong buffer scheduling to hide dispatch latency, and the VkMemoryHeap budget that prevents OOM on Adreno 750