opencl: stop writing zeros into the padded MoE activation slots

kernel_moe_reorder_b zero-filled every padded token slot. The MoE GEMMs
accumulate per output column and scatter only the real columns, so whatever sits
in a padded slot never reaches dst and the fill is pure bandwidth.

Established by a poison test rather than by reading the closed-source kernel:
filling padded slots with 1e30 leaves MUL_MAT_ID at 383 OK / 0 FAIL. The
positive control is what gives that meaning -- poisoning the real gather as well
fails 188 of 383 with NaN at index 0, and 188 is exactly the moe_reorder_b
dispatch count, so the buffer is read and the test can fail.

The destination is the reused prealloc_act_trans scratch, so padded slots now
retain stale activations from a previous layer rather than zeros. That is safe
for the same reason.

Worth about 5% of the kernel on granite-3b MXFP4. Perplexity is bit-identical on
three models, which covers the cross-layer stale-buffer case test-backend-ops
cannot reproduce.

Author: wanghqc

ggml/src/ggml-opencl/kernels/moe_reorder_b.cl

@@ -20,11 +20,15 @@ kernel void kernel_moe_reorder_b(
 
     uint router_idx = router[post_router_idx];
 
-    float4 out = (float4)(0);
-    if (router_idx != 0xFFFFFFFF) {
-        ushort activation_idx = router_idx / map_ratio;
-        out = src[activation_idx * K / 4 + k_4];
+    // Padded slots need not be written at all. The MoE GEMMs accumulate per output
+    // column and scatter only the real columns, so whatever sits in a padded slot
+    // never reaches dst -- verified by filling them with 1e30 and re-running
+    // test-backend-ops MUL_MAT_ID (383 OK / 0 FAIL, unchanged), against a positive
+    // control that poisons the real gather too and fails 188 of 383.
+    if (router_idx == 0xFFFFFFFF) {
+        return;
     }
 
-    dst[post_router_idx * K / 4 + k_4] = out;
+    ushort activation_idx = router_idx / map_ratio;
+    dst[post_router_idx * K / 4 + k_4] = src[activation_idx * K / 4 + k_4];
 }