Add vecadd-loop

This is the same kernel as vecadd but repeated in a for-loop many times so that
the runtime overhead at the startup is amortized.

tests/opencl/vecadd-loop/kernel.cl.loop

@@ -0,0 +1,9 @@
+__kernel void vecadd_loop (__global volatile const float *A,
+                    __global volatile const float *B,
+                    __global volatile float *C)
+{
+  int gid = get_global_id(0);
+  for (int i = 0; i < 100; i++) {
+    C[gid] = A[gid] + B[gid];
+  }
+}