Seed the non qjit reference run of a test (#1757)

**Context:**
The test 
`pytest/test_measurement_transforms.py::test_measurement_from_samples_with_sample_measurement`
has a [non-qjit reference
run](https://github.com/PennyLaneAI/catalyst/blob/02af052424ec4fd5ce7d4e846c3410d823e13c55/frontend/test/pytest/test_measurement_transforms.py#L445)
that is unseeded and is causing some flaky test noise.

**Description of the Change:**
Since lightning.qubit does not support seeding, we use default.qubit for
this reference run and seed it.

**Benefits:**
No more flaky CI failures on this test.

**Possible Drawbacks:**
I don't like using seeding to silence flaky tests. 
In our [guidance docs for flaky
tests](https://github.com/PennyLaneAI/guidance-docs/blob/eae480fd82f88c5f922e4d1b5a324819a4690b8a/development/testing-guidelines.md?plain=1#L222),
seeding is only the second suggested strategy.
But adopting the first strategy of hypothesis testing would take more
time. For now, the band-aid seed solution would have to do.

Author: paul0403

frontend/test/pytest/test_measurement_transforms.py

@@ -438,11 +438,15 @@ def circuit(theta: float):
 
         theta = 2.5
         res = qjit(measurements_from_samples(circuit, dev.wires), seed=37)(theta)
-
+        # PL flattens N-by-1 2D result arrays into size-N 1D arrays, but Catalyst does not
         if len(measurement().wires) == 1:
-            samples_expected = qjit(circuit, seed=37)(theta)
-        else:
-            samples_expected = circuit(theta)
+            res = res.flatten()
+
+        # lightning.qubit does not support seeding.
+        # To resolve flakiness, we put the non qjit reference run on default.qubit,
+        # which can be seeded
+        ref_dev = qml.device("default.qubit", wires=4, shots=3000, seed=42)
+        samples_expected = qml.qnode(ref_dev)(circuit.func)(theta)
 
         assert res.shape == samples_expected.shape
         assert np.allclose(np.mean(res, axis=0), np.mean(samples_expected, axis=0), atol=0.05)