Generalized plot_posterior to plot vector-valued variables

[fonnesbeck]

This is a refactor of plot_posterior that accepts vector-valued parameters. I will add a test, but please test it on your favorite example.

Closes #1131

[ColCarroll]

Cool! Works on occupancy example -- changing n=100 to n = (2, 100) in the setUp creates 200 plots for z, which seems correct.

Maybe a warning for too many plots? Or an implicit upper bound?

[fonnesbeck]

We don't warn users for other plots, so I wouldn't put one here. If we think that's something we ought to have then it's probably best implemented module-wide in a separate PR.

[springcoil]

I see nothing wrong with this fix. Maybe @ColCarroll has some objections though.

[ColCarroll]

No, this looks good -- @fonnesbeck should we wait for a test or merge now? I was using the test below, if that helps.

import pymc3 as pm
import numpy as np

def test_plot_posterior(self):
    n = (2, 100)
    theta = 2.1
    pi = 0.4
    y = (np.random.random(n) < pi) * np.random.poisson(lam=theta, size=n)

    with pm.Model() as model:
        psi = pm.Beta('psi', 1, 1)
        pm.Bernoulli('z', psi, y.shape)
        theta = pm.Uniform('theta', 0, 100)
        pm.ZeroInflatedPoisson('y', theta, psi, observed=y)
        start = {'psi': 0.5, 'z': (y > 0).astype(int), 'theta': 5}
        step_one = pm.Metropolis([model.theta_interval_, model.psi_logodds_])
        step_two = pm.BinaryMetropolis([model.z])
        tr = pm.sample(50, [step_one, step_two], start)

    # should produce a plot for each dimension of n
    plots = pm.plot_posterior(tr, varnames=['z'])
    self.assertEqual(plots.shape, (np.prod(n),))

[twiecki]

I think we should add @ColCarroll's test.

[fonnesbeck]

The added test should do it, and it uses an existing test model run.

[ColCarroll]

Looks good to me -- will merge after travis passes