Further Improve robustness of model classifier tests

[chrishalcrow]

After merging (#4447), there were still flakey failures in the tests.

This PRs makes the training data more robust in two ways:

1. The number of units in the generated recording has been reduces. This is better because the units are spaced out with some minimum distance between each one. If the number is too high, a unit can be produced quite far from the probe and hence have a small amplitude.
2. We give the model repeated data, ensuring it sees all the data in its training. Hence it really should not make a mistake. You don't want to do this when actually training a model, but good for testing.

When testing before, I used pytests count flag. This was stupid: the test was using a cache so wasn't retrained each time. This time I ran the test repeatedly from scratch using something like

for i in {1..1000}; do echo "--- Run #$i ---"; pytest test_file.py::test_name -q || break; rm -r path_to_cache; done

Thanks Gemini!

No failures on 1000 runs :)