[CUDA] Bump tolerances in test_svd_lowrank_cuda_float64 (#143049)

pre-emptive bump for apparent noisy failure

Pull Request resolved: https://github.com/pytorch/pytorch/pull/143049
Approved by: https://github.com/Skylion007, https://github.com/lezcano, https://github.com/nikitaved

test/test_linalg.py

@@ -2506,7 +2506,7 @@ class TestLinalg(TestCase):
             # check if svd_lowrank produces same singular values as linalg.svdvals
             U, S, Vh = torch.linalg.svd(a, full_matrices=False)
             V = Vh.mH
-            self.assertEqual(s, S)
+            self.assertEqual(s, S, rtol=5e-7, atol=1e-7)
 
             if density == 1:
                 # actual_rank is known only for dense inputs