feat: fashion mnist and non robustness filtering

[flekschas-ozette]

This PR fixes one small bug (when the categorical labels are not of type string) and filters out non-robust points from the metric scatter. The later allows the user to keep non-robust points in the categorical scatter (drawn as dark gray points) without being visually impacted by their metric value.

Most importantly, this PR adds a useful for how to compare different embedding methods (with and without existing label information) using Fashion MNIST data.

[flekschas]

@manzt shall we still merge this? The string check for checking whether the label is a marker could still be useful I think

[manzt]

Good question! I think the changes to the src are good, but do you you still want to add the notebook?

[manzt]

Actually, wasn't this implemented in #43?

[flekschas-ozette]

Not this part:

self.has_markers = (
    isinstance(self._data[_LABEL_COLUMN][0], str)
    and "+" in self._data[_LABEL_COLUMN][0]
)

We currently assume that the label column is a string but it doesn't have to be. Ints can also define categories (like in the clustering case where the labels are just cluster ints)

I think having this notebook could be useful in general? Like when someone finds this tool but doesn't have labels yet? What do you think?