dflash: enable Qwen3-Coder-Next on Vulkan

[gboddaer]

Summary

Enable Qwen3-Coder-Next DFlash on Vulkan by fixing the runtime path that caused 0% draft acceptance when the GPU cross ring is unavailable.

Key changes:

• Add dflash_target_kv_available to the drafter context parameters.
• Set that flag from DFlash setup based on whether the GPU cross ring exists.
• Gate full-attention DFlash KV-cache reads on that flag so Vulkan CPU-hidden capture falls back to fresh K/V projection from target_hidden instead of reading empty/stale target KV.
• Add Qwen3Next pre-conv QKV tape aliases so DFlash conv-state replay advances r_l correctly.
• Keep flat/batched DFlash drafting on the full block_size query block while consuming only n_draft + 1 output rows.
• Add focused plumbing tests and a concise rationale document.

Verification

Ran in isolated worktree based on latest origin/main:

CPU CMake configure: rc=0
CPU test-dflash-plumbing target build: rc=0
build_cpu/bin/test-dflash-plumbing: rc=0
CPU full build: rc=0

Vulkan configure was re-tested after installing glslc locally on Debian 11:

/home/gbo/.local/bin/glslc --version: shaderc v2023.2
cmake -B build_vulkan -DGGML_VULKAN=ON ...: rc=0
Found Vulkan: /usr/lib/x86_64-linux-gnu/libvulkan.so (found version "1.2.162") found components: glslc

Vulkan compilation then failed later because the Debian 11 Vulkan headers are too old for current ggml-vulkan.cpp symbols, for example:

vk::PhysicalDeviceMaintenance4Properties is not a member of vk
vk::DriverId::eMesaTurnip is not a member of vk::DriverId

So the original missing-glslc configure blocker is resolved on this host; full Vulkan compilation requires newer Vulkan headers/SDK than Debian 11's vulkan.hpp 1.2.162.

Notes

This PR intentionally excludes drafter training scripts and broad investigation artifacts. It includes one canonical rationale doc:

• docs/enable-dflash-qwen3-coder-next-on-vulkan.md