[AMD] Fix addresses for other stores from BufferLoadToLocal (#7016)

We need to predicate the `llStore` for the other value based on the
initial mask or otherwise we would need to shuffle the other value as
well.

`AsyncCopyGlobalToLocal` does this already correctly.

Author: AlexAUT

third_party/amd/lib/TritonAMDGPUToLLVM/LoadStoreOpToLLVM.cpp

@@ -561,8 +561,8 @@ struct BufferLoadToLocalOpConversion
             rewriter, this->getTypeConverter(), loc, vecTy, otherElems, srcIdx);
         llStore(rewriter, loc,
                 hasSwizzling ? swizzledShmemAddr[i] : coalescedShmemAddr[i],
-                storeVal, b.icmp_ne(pred, b.true_val()), op.getCache(),
-                /*forceNoAliasAsyncLoads=*/true);
+                storeVal, b.icmp_ne(maskElems[srcIdx], b.true_val()),
+                op.getCache(), /*forceNoAliasAsyncLoads=*/true);
       }
     }