vulkan: Switch MUL_MAT_VEC to 4 K per iteration for F16/32 (#22887)

* vulkan: Switch MUL_MAT_VEC to 4 K per iteration for F16/32

Against mesa git, this shows a 4.8% performance improvement for
tg128 on Qwen3.5-9B:BF16 on Intel BMG.

Note that this breaks some tests until the last commit which fixes
OOB A reads.

* vulkan: Use aligned loads in mul_mat_vec when available

Against mesa git, this shows a 3.3% performance improvement for
tg128 on Qwen3.5-9B:BF16 on Intel BMG.

* Make explicit that `num_rows` is <= `NUM_ROWS` in mul_mat_vec

Mesa's UUB logic can't see through conditionals, limiting its
ability to understand the bounds on the `num_rows` field in the
cleanup run. Making it explicit that `num_rows` is, indeed, always
<= `NUM_ROWS` helps mesa make slightly better codegen.

Against mesa git, this currently shows a 1% performance improvement
in tg128 on Qwen3.5-9B:BF16 on Intel BMG.

* vulkan: Fix OOB A reads in MUL_MAT_VEC for odd sizes

There was a TODO to fix the OOB reads from the A matrix which we do
here.

It is within performance noise (+<0.1%) in tg128 for
Qwen3.5-9B:BF16 on Intel BMG.

ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl

@@ -5,21 +5,60 @@
 #include "types.glsl"
 
 #if defined(DATA_A_F32)
+FLOAT_TYPE dequantize1(uint ib, uint iqs, uint a_offset) {
+    return data_a[a_offset + ib];
+}
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     return vec2(data_a[a_offset + ib], data_a[a_offset + ib + 1]);
 }
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    return vec4(data_a[a_offset + ib    ], data_a[a_offset + ib + 1],
+                data_a[a_offset + ib + 2], data_a[a_offset + ib + 3]);
+}
+vec4 dequantize4_2aligned(uint ib, uint iqs, uint a_offset) {
+    return vec4(data_a[a_offset + ib    ], data_a[a_offset + ib + 1],
+                data_a[a_offset + ib + 2], data_a[a_offset + ib + 3]);
+}
+
 #endif
 
 #if defined(DATA_A_F16)
+FLOAT_TYPE dequantize1(uint ib, uint iqs, uint a_offset) {
+    return data_a[a_offset + ib];
+}
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     return vec2(data_a[a_offset + ib], data_a[a_offset + ib + 1]);
 }
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    return vec4(data_a[a_offset + ib    ], data_a[a_offset + ib + 1],
+                data_a[a_offset + ib + 2], data_a[a_offset + ib + 3]);
+}
+vec4 dequantize4_2aligned(uint ib, uint iqs, uint a_offset) {
+    const vec2 a = data_a_packed32[(a_offset + ib)/2];
+    const vec2 b = data_a_packed32[(a_offset + ib)/2 + 1];
+    return vec4(a, b);
+}
 #endif
 
 #if defined(DATA_A_BF16)
+FLOAT_TYPE dequantize1(uint ib, uint iqs, uint a_offset) {
+    return bf16_to_fp32(data_a[a_offset + ib]);
+}
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     return vec2(bf16_to_fp32(data_a[a_offset + ib]), bf16_to_fp32(data_a[a_offset + ib + 1]));
 }
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    return vec4(bf16_to_fp32(data_a[a_offset + ib    ]), bf16_to_fp32(data_a[a_offset + ib + 1]),
+                bf16_to_fp32(data_a[a_offset + ib + 2]), bf16_to_fp32(data_a[a_offset + ib + 3]));
+}
+vec4 dequantize4_2aligned(uint ib, uint iqs, uint a_offset) {
+    const uint a = data_a_packed32[(a_offset + ib)/2];
+    const uint b = data_a_packed32[(a_offset + ib)/2 + 1];
+    return vec4(uintBitsToFloat((a & 0x0000ffff) << 16),
+                uintBitsToFloat( a & 0xffff0000),
+                uintBitsToFloat((b & 0x0000ffff) << 16),
+                uintBitsToFloat( b & 0xffff0000));
+}
 #endif
 
 #if defined(DATA_A_Q4_0)