Arviz: Change pair_plot behavior for MCMC methods that do not produce divergences

[canyon289]

Describe the bug
When using metropolis hasting sampler in pymc3, then converting to inference data, then using az.pair_plot(divergences=True) raises an index error

To Reproduce

import matplotlib.pyplot as plt
import arviz as az
import pymc3 as pm

x = [1,2]
y = [1,2]
with pm.Model() as model_g:
    α = pm.Normal('α', mu=0, sd=10)
    β = pm.Normal('β', mu=0, sd=1)
    ϵ = pm.HalfCauchy('ϵ', 5)

    μ = pm.Deterministic('μ', α + β * x)
    y_pred = pm.Normal('y_pred', mu=μ, sd=ϵ, observed=y)

# NUTS sampler works when plotting divergences
"""
with model_g:
    trace_nuts_non_centered = pm.sample()

nuts_non_centered_dataset = az.from_pymc3(trace=trace_nuts_non_centered)
az.plot_pair(nuts_non_centered_dataset, var_names=['α', 'β', 'ϵ'], divergences=True)

plt.show()
"""

# Metropolis Hastings raises indexerror
with model_g:
    step = pm.Metropolis()
    trace_mh_non_centered = pm.sample(step=step)

az.plot_pair(trace_mh_non_centered, var_names=['α', 'β', 'ϵ'], divergences=True)
plt.show()

Expected behavior
Divergences are plotted as expected
Show a warning and continue plotting

Additional context
Versions of arviz and other libraries used, operating system used, and anything else that may be useful.

Arviz Version '0.3.3'
PyMC3 Version : '3.6'

[ahartikainen]

So it should raise some other error? Or plot nothing?

[junpenglao]

that's because MH does not produce divergences information in diagnostic right? what about setting divergences=False

[aloctavodia]

Divergences are False by default. If divergence information is not present and divergences=True we should make the plot anyway and show a warning.

[canyon289]

Divergences are False by default. If divergence information is not present and divergences=True we should make the plot anyway and show a warning.

Ah shoot yea i wrote the solution up there wrong. I think we should what @aloctavodia suggested. I can make a PR