Relax tolerance of UnitaryOverlap tests

The failure rate for the parametric identity tests was about 0.2%, which
at our CI scale corresponds to approximately one spurious failure every
two days.  The alternative to this is to fix the random seed of the
test, but similar to other `quantum_info` tests, we would like some
degree of extended coverage.  Lifting the tolerance by three orders of
magnitude at this circuit size should ensure a zero false-positive rate.

This also changes the test to actually display the failure tolerance, so
it's easier to tell from a CI run if a failure was real.

test/python/circuit/library/test_overlap.py

@@ -30,7 +30,7 @@ def test_identity(self):
         unitary.assign_parameters(np.random.random(size=unitary.num_parameters), inplace=True)
 
         overlap = UnitaryOverlap(unitary, unitary)
-        self.assertTrue(abs(Statevector.from_instruction(overlap)[0] - 1) < 1e-15)
+        self.assertLess(abs(Statevector.from_instruction(overlap)[0] - 1), 1e-12)
 
     def test_parameterized_identity(self):
         """Test identity is returned"""
@@ -40,7 +40,7 @@ def test_parameterized_identity(self):
         rands = np.random.random(size=unitary.num_parameters)
         double_rands = np.hstack((rands, rands))
         overlap.assign_parameters(double_rands, inplace=True)
-        self.assertTrue(abs(Statevector.from_instruction(overlap)[0] - 1) < 1e-15)
+        self.assertLess(abs(Statevector.from_instruction(overlap)[0] - 1), 1e-12)
 
     def test_two_parameterized_inputs(self):
         """Test two parameterized inputs"""