Restore approx 0 rounding to decomp0 method

In Qiskit#4656 we added rounding to the output of the decomp0 method to handle
a case where differing FP precision on windows environments was causing
an expected result in running two_qubit_cnot_decompose on np.eye(4)
with numpy 1.19.x installed leading to a hard failure in the qasm tests.
This seemed to reliably unblock testing and make unit tests work
reliably. However, that original fix from Qiskit#4656 was superseded by Qiskit#4835
which was a fix for a more general issue with the reproducibility of the
decompositions and reverted. Since Qiskit#4835 has merged we've been seeing an
uptick in the failure rate on the same unitary qasm test that Qiskit#4656
fixed, so the change in Qiskit#4835 was not actually sufficient for the
windows case. This commit restores the fix from Qiskit#4656 to unblock CI and
fix the reproducability of the decompositions across systems.

Fixes Qiskit#4856

qiskit/quantum_info/synthesis/two_qubit_decompose.py

@@ -382,15 +382,18 @@ def traces(self, target):
                 4]
 
     @staticmethod
-    def decomp0(target):
+    def decomp0(target, eps=1e-15):
         """Decompose target ~Ud(x, y, z) with 0 uses of the basis gate.
         Result Ur has trace:
         :math:`|Tr(Ur.Utarget^dag)| = 4|(cos(x)cos(y)cos(z)+ j sin(x)sin(y)sin(z)|`,
         which is optimal for all targets and bases"""
 
         U0l = target.K1l.dot(target.K2l)
         U0r = target.K1r.dot(target.K2r)
-
+        U0l.real[abs(U0l.real) < eps] = 0.0
+        U0l.imag[abs(U0l.imag) < eps] = 0.0
+        U0r.real[abs(U0r.real) < eps] = 0.0
+        U0r.imag[abs(U0r.imag) < eps] = 0.0
         return U0r, U0l
 
     def decomp1(self, target):