vulkan: Support GGML_TYPE_NVFP4 (#21455)

This adds nvfp4 support for get_rows, dequant, and mul_mat(_id). For
mul_mat, it does not add support for the dp4/q8_1 path, it's all via
fp16/fp32.

ggml/src/ggml-vulkan/vulkan-shaders/dequant_nvfp4.comp

@@ -0,0 +1,32 @@
+#version 450
+
+#include "dequant_head.glsl"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_nvfp4 data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    const uint tid = gl_LocalInvocationID.x % 64;
+    const uint sub = tid / 16;
+    const uint ir = tid % 16;
+    const uint ib = 16 * i + ir;
+    if (ib >= p.nel / 64) {
+        return;
+    }
+
+    const uint q_idx = 8 * sub;
+    const uint b_idx = 1024 * i + 64 * ir + 16 * sub;
+
+    const float d = ue4m3_to_fp32(data_a[ib].d[sub]);
+
+    [[unroll]] for (uint l = 0; l < 8; ++l) {
+        data_b[b_idx + l +  0] = D_TYPE(d * 0.5 * float(kvalues_mxfp4[data_a[ib].qs[q_idx + l] & 0xF]));
+        data_b[b_idx + l +  8] = D_TYPE(d * 0.5 * float(kvalues_mxfp4[data_a[ib].qs[q_idx + l] >>  4]));
+    }
+}