Added support for Vulkan, any interest?

I have a Intel Arc A770 16GB GPU and wanted to use it with candle.
I took niklasha's work on niklas-vulkan-2 branch cherry-pick's into the current main branch.
I (when I say I, I mean I was the navigator, Codex 5.2 max did the work) added the following:

Added Vulkan queue-family selection and synchronize() so VulkanDevice uses compute-capable queues and can block on GPU work (device.rs).
Expanded Vulkan storage surface with raw_buffer() access for kernel dispatch and fixed error wiring (storage.rs).
Wired Vulkan kernel registry to include matmul, norms, softmax, masked softmax, and quantized kernels (lib.rs).
Added F32/F16 matmul shader stubs and norm/softmax shaders for initial Vulkan ops coverage (*.comp).
Implemented Vulkan masked softmax and staged SDPA path with GQA support in candle-nn (ops.rs).
Added Vulkan smoke tests and masked softmax correctness test (vulkan_smoke_tests.rs, vulkan_masked_softmax.rs).
Fixed missing imports and push-constant binding for Vulkan command execution (storage.rs).
Added bytemuck + vulkano-shaders feature wiring for Vulkan builds (Cargo.toml).
Introduced QVulkanStorage backed by raw byte buffers with dequantize/quantize helpers (vulkan.rs).
Added Vulkan quantized matmul kernels for Q5_0 and Q8_0 (naive, F32 output) (qmatmul_q5_0_f32.comp, qmatmul_q8_0_f32.comp).
Hooked Vulkan quantized path into QTensor forward and added Vulkan quantized tests (mod.rs, vulkan_quantized_tests.rs).
Added a dequantize‑fallback backward path for QLoRA-style gradients (mod.rs).
Cleaned up dummy Vulkan stubs to match new quantized API surface (dummy_vulkan.rs).
Fixed multiple test harness macro/feature mismatches to compile with Vulkan enabled (test_utils.rs, *.rs).

Review the Vulkan changes listed across device.rs, storage.rs, lib.rs, ops.rs, vulkan.rs, mod.rs, and the named shader and test files to understand the proposed scope. Run the Vulkan smoke, masked-softmax, and quantized tests with Vulkan enabled, then determine whether this broad implementation is wanted and what maintainers consider complete.

Written by the indexing model from the issue text.