Fix IPU Jacobi eigh algorithm when size % 4 == 2

The recent improvement in PR #49 introduced a regression in Jacobi
`eigh`, raising an error when size % 4 == 2. This is due to the partial
loop unrolling in Jacobi second update stage.

This PR is fixing the issue by passing explicitely the offset and size
of the workload to the vertex.

tests/linalg/test_tile_linalg_jacobi.py

@@ -177,8 +177,11 @@ def test__jacobi_eigh__single_iteration(self):
         npt.assert_array_almost_equal(np.linalg.eigh(A)[0], np.linalg.eigh(x)[0], decimal=5)
 
     @unittest.skipUnless(ipu_num_tiles >= 16, "Requires IPU with 16 tiles")
-    def test__jacobi_eigh_raw__proper_eigh_result(self):
-        N = 12
+    @parameterized.parameters(
+        {"N": 10},  # testing Jacobi 2nd update.
+        # {"N": 12},
+    )
+    def test__jacobi_eigh_raw__proper_eigh_result(self, N):
         x = np.random.randn(N, N).astype(np.float32)
         x = (x + x.T) / 2.0