autotest: improve test.Plane.CompassLearnInFlight robustness

The test has always learned a marginal compass value (25-30, with
failure at 35) and since commit `a52f911b` it has usually failed on my
machine with a value of 40-45 presumably due to a small change in the
flight path. It also occasionally fails in CI with similar values;
presumably different machine speeds select a different part of the
flight during which learning occurs.

Fix the issue by flying a straight level path to allow time for the GSF
to converge, as that is used as the reference for learning according to
the code. This results in a learned value of 5-10 which is much closer
to the target and hopefully less sensitive to the particular machine.

Author: tpwrules

Tools/autotest/arduplane.py

@@ -6737,6 +6737,9 @@ def CompassLearnInFlight(self):
         self.wait_ready_to_arm()
         self.takeoff(30, mode='TAKEOFF')
         self.assert_parameter_value("COMPASS_OFS_X", 20, epsilon=30)
+        # fly straight and level for a bit to let GSF converge for accurate learning
+        self.change_mode("FBWB") # not "CRUISE" to avoid heading track with bad compass
+        self.wait_distance(200, accuracy=20)
         old_compass_ofs_x = self.get_parameter('COMPASS_OFS_X')
         self.set_parameters({
             "COMPASS_OFS_X": 1100,