AP_Compass: throw away all samples which aren't acceptable in thinning

[peterbarker]

From my 5-year-old notes:

+Getting a good list of samples for the fits to run across is probably a good thing to get working well.
+
+I don't think we're doing that at the moment.
+
+At the moment: we test a sample, and if it is no good then we copy a random one in from the end of the list.  We do not test the sample we just copied in, rather move on through the list.
+
+This PR changes things such that we test every sample in the buffer as part of the thinning step.
+
+```
+2019-12-06 10:07:40.92: STATUSTEXT {severity : 6, text : Threw away 247/300 samples}
+2019-12-06 10:07:41.54: STATUSTEXT {severity : 6, text : Threw away 232/300 samples}
+```
+
+There are 576 faces on the default polyhedron.   This test shows that we're only hitting 50-70 of them in the initial collect-samples pass.

+After the initial collect samples pass we fill the buffer up with completely untested samples - I can't imagine we'd do better than the same again.  So overall we'll be hitting 20% of the faces with no guarantee on face distribution, except inasmuch

I have no idea how I was going to finish that sentence.

The "threw away line was extra debug in the thinning_samples method recording before/after sample counts.

I've tested this on a Durandal quadcopter here and the calibration result is similar to what was on there.

[andyp1per]

Couldn't this get to 0 meaning the check above would be a free memory read?

[peterbarker]

Then i would equal samples collected, so we would exit the while loop

[tridge]

maybe i < _samples_collected-1 ? otherwise copies to itself

[tridge]

i think this looks good, but should have a bit more testing