vulkan: add GATED_DELTA_NET op support

[ProgenyAlpha]

Summary

Adds Vulkan compute shader implementation for the fused GATED_DELTA_NET op (merged in #19504), enabling the recurrence fusion for Vulkan users running delta net models (Qwen3.5, Qwen3-Coder-Next, etc).

• Fused compute shader (gated_delta_net.comp) — one workgroup per (head, sequence), one thread per state column
• Specialization constants for head size (32/64/128) and KDA mode — six pipeline variants compiled from one shader source
• Full support: scalar gate, KDA vector gate, GQA broadcast (v_repeat), multi-token, permuted non-contiguous q/k
• Matches the CUDA kernel's fused decay + rank-1 update pattern (S[i] = g * S[i] + k[i] * delta)
• Uses beta strides (sb*) consistent with the merged op interface

Status

WIP / Draft — still needs real model benchmarking and multi-vendor driver testing.

• 13/13 test-backend-ops passed (AMD Radeon 890M, RADV GFX1150)
• Real model benchmarking (Qwen3.5 inference)
• Testing on NVIDIA and Intel Vulkan drivers
• Cache exp() in KDA path (currently computed twice per loop)
• Verify local_size_x_id spec constant behavior across drivers

Mentions #19504, #14909