plot_ppc only plots one legend if there are multiple observed

[MarcoGorelli]

To Reproduce

import seaborn as sns
import pymc3 as pm
import pandas as pd
import arviz as az

iris = sns.load_dataset("iris")
y_s = pd.Categorical(iris["species"]).codes
x_n = "sepal_length"
x_0 = iris[x_n].values

with pm.Model() as lda:
    μ = pm.Normal("μ", mu=0, sd=10, shape=2)
    σ = pm.HalfNormal("σ", 10)
    setosa = pm.Normal("setosa", mu=μ[0], sd=σ, observed=x_0[:50])
    versicolor = pm.Normal("versicolor", mu=μ[1], sd=σ, observed=x_0[50:])
    bd = pm.Deterministic("bd", (μ[0] + μ[1]) / 2)
    trace_lda = pm.sample(1000, return_inferencedata=True)

with lda:
    ppc = pm.sample_posterior_predictive(trace=trace_lda)

az.plot_ppc(az.from_pymc3(posterior_predictive=ppc, model=lda))

Expected behavior
I think I'd have expected both subplots to have a legend

Additional context

> az.__version__, pm.__version__
('0.11.1', '3.11.0')

[aloctavodia]

We can change it to optionally have more than one legend, but I think this default makes sense at it reduces redundancy (and the data/ink ratio).

[MarcoGorelli]

Sure, but the second subplot is about versicolor (whilst the first one setosa), so is it really redundant to have an extra legend when the first one reads "observed setosa"?

[aloctavodia]

Yes, if you have that information on the x_label. Of course the current behavior is problematic, because we have one legend that is supposed to work for both plots but is referring only to the first. So I suggest to have as default one legend and remove from the legend the name of the variable.

[MarcoGorelli]

Agreed, that would be good - I'll work on a PR next week

[aloctavodia]

Thanks @MarcoGorelli!