vulkan: use mul_mat_vec_id for high-expert-count MoE decode

[essentialols]

Extends ggml_vk_use_mul_mat_vec_id() to route MoE models with a large number of experts (>64, e.g. DeepSeek-V3/V4-class 256-expert models) to the mul_mat_vec_id path in the decode regime (batch <= 256).

Motivation

On RADV (Mesa; tested on an AMD Radeon 680M / gfx1035, Mesa 26.1.4) the mul_mat_id shader for >=256-expert MoE models stalls during pipeline compilation at high --n-gpu-layers, wedging the device (SMU becomes unresponsive: SMU: No response, Ring sdma0 reset failed, GPU reset ... ret=-62). Routing these dispatches to the existing mul_mat_vec_id path avoids the stall and is also faster per-dispatch for token generation.

Change

ggml_vk_use_mul_mat_vec_id(): alongside the existing small-batch case (batch <= 8), also return true when src0->ne[2] > 64 (>64 experts) and src2->ne[1] <= 256, restricting the change to the decode regime so large prefill batches are unaffected.

Testing

AMD Radeon 680M (gfx1035), Vulkan/RADV, Mesa 26.1.4, build b10066.

• Without the change: a 256-expert MoE (Qwen3.6-35B-A3B-class) hangs during shader compilation at -ngl 99 (empty output, device wedge requiring a GPU reset).
• With the change: the same model loads and generates coherently at -ngl 99.
• No behavioral change for models with <=64 experts or for batch > 256.

[ggml-gh-bot]

Hi @essentialols, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• PR Template not respected: Please respect the template when creating a new pull request. Make sure to fill out all required sections.

• Multiple open PRs from a new contributor: We limit new contributors (those without a previously merged PR) to 1 open PR at a time. You currently have 3 open PRs.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.

[0cc4m]

Open an issue and try to identify why it hangs first. Most likely the validation layers will show some problem.