Vulkan accelerator backend (Device::Vulkan): implementation available, seeking to upstream #3985

Summary

Proposing a Vulkan backend for candle-core analogous to the existing CUDA and Metal backends. A working reference implementation exists in the rexlunae/joshua fork. This issue is to gauge interest and agree a shape before upstreaming a ~2k-line backend.

What the implementation provides

• Device::Vulkan / DeviceLocation::Vulkan + VulkanDevice / VulkanStorage, wired through the Device/Storage/DeviceLocation enums and all allocation/sync paths — mirrors the Metal/OpenCL-style backend structure.
• Uses ash with its loaded feature (libvulkan.so dlopen'd at runtime, no hard link) and naga to compile embedded GLSL compute shaders to SPIR-V at runtime (no external glslc).
• Bring-up target is unified-memory iGPUs (AMD Renoir / RADV): host-visible+coherent buffers make the host<->device round-trip a plain memcpy; native kernels cover matmul (register-blocked, handles transposed/batched/broadcast rhs), elementwise, and last-dim reduction.
• CustomOp1/CustomOp2::vulkan_fwd device-native hooks so softmax/RMSNorm-style CustomOps can run on-device instead of the CPU default (default remains a correct CPU round-trip).

Validation (AMD Radeon / RADV RENOIR iGPU)

Native kernels are bit-exact vs CPU (matmul, elementwise, last-dim reduction). Device-native softmax matches CPU to ~7.5e-9 and RMSNorm to ~1e-7. Full 150B-MoE inference runs on-device via --device vulkan.

Questions

1. Is there interest in a Vulkan backend upstream? It parallels the Metal backend's role (broader hardware coverage than CUDA-only), especially for iGPUs and RADV-based systems.
2. Shape preference: full backend in one PR, or a staged series (device/storage round-trip first, then kernels, then CustomOp hooks)?
3. Any concerns about the ash/naga dependencies (currently optional feature-gated behind vulkan)?

I'm happy to prepare the PR in the agreed shape using the joshua implementation as the reference.