llama.cpp Vulkan split offload hits OOM on B580

This is a common pain point with Vulkan backends — memory management is less mature than CUDA, and split GPU/CPU offloading is notoriously finicky on non-NVIDIA hardware.

• –The `--fit` flag attempts auto-allocation but can overcommit VRAM on Intel Arc, triggering Vulkan device loss
• –`--no-mmap` forces the full model into RAM before layer splitting, which is the correct approach but requires careful `-ngl` tuning to avoid VRAM overflow
• –The correct approach is explicit `-ngl <n>` to control exact layer count on GPU, combined with `--no-mmap` and sufficient RAM headroom
• –Intel Arc B580 has 12GB VRAM; an 80B Q3_K_M model is ~35GB, so only a fraction of layers fit on GPU
• –Vulkan support in llama.cpp continues to lag behind CUDA for advanced memory features like unified addressing