Fix KeyError when using column of pd.Categorical dtype with unobserved categories

[arcanaxion]

This PR passes observed=True to omit categorical values with no observations in the dataset. Fixes #4274 and #4433.

Also silences the FutureWarning from pandas-dev/pandas#43999 since pandas 2.1 that the default of observed=False will be changed to True in a future version.

Following code raises KeyError: 'Fri' before this fix:

import plotly.express as px
df = px.data.tips()
df["day"] = df["day"].astype("category")
px.scatter(
    df[df.day != "Fri"],
    x="total_bill",
    y="tip",
    facet_col="day",
)

[alexcjohnson]

💃 Thanks @arcanaxion, your fix looks great, and thanks so much for the clean & clear test.

CI failure looks flaky, I've rerun it. Would you add a changelog entry, then we can merge?

[arcanaxion]

Thanks @alexcjohnson! Just added a changelog entry. Will this do?

[Coding-with-Adam]

Thank you for this PR @arcanaxion
The community's support in building new features or fixing bugs is of great importance and very much appreciated. If there is anything else that catches your eye, let us know.