fix(dsv4): lightning-indexer bounds (as built into b10867 from a dirty tree)

Author: Stratum

ggml/src/ggml-cuda/lightning-indexer.cu

@@ -327,7 +327,11 @@ static __global__ void lightning_indexer_kernel_vec(
 
         constexpr int n_q = N_HEAD_INNER * (N_EMBD / 4);
 #pragma unroll
-        for (int i_q = tid; i_q < n_q; i_q += THREADS_PER_BLOCK) {
+        for (int i_q0 = 0; i_q0 < n_q; i_q0 += THREADS_PER_BLOCK) {
+            const int i_q = i_q0 + tid;
+            if (i_q >= n_q) {
+                break;
+            }
             const int i_head_inner = i_q / (N_EMBD / 4);
             const int i_head = i_head_0 + i_head_inner;
             const int i_embd = i_q % (N_EMBD / 4);