+1 to @ricardoV94, it is difficult to construct a simple example with multimodal posterior. The posterior usually easily overwhelm by the likelihood (given that it is a simple model), on top of that if your likelihood is not multimodal (which the case when you are using Normal) you wont see multimodal in the posterior.
Not using Gaussian, with few data point usually makes it a bit easier to construct multimodal posterior:
with pm.Model():
b = pm.StudentT("b", nu=4., mu=-5., sigma=1.)
obs = pm.StudentT("observed", nu=4., mu=b, sigma=1., observed=5.)
idata = pm.sample()
az.plot_trace(idata);
