fix brittle test loo by modifying test dataset

[Cab14bacc]

closes #103

The original my_data dataset:

We modify by making y the clearly better predictor:

 data["z"] = data["y"] + 0.1 * np.random.normal(size=len(data))

Tested with this:

NUM_DRAWS, NUM_CHAINS = 100, 4

df_cpy = bmb.load_data("my_data") 
df_cpy["z"] = 2.0 * df_cpy["y"] + np.random.normal(0, scale=0.1, size=len(df_cpy))
data = df_cpy.copy()

result = []
model = bmb.Model("z ~ x + y", data)  

for i in range(100):
    idata = model.fit(
        draws=NUM_DRAWS, chains=NUM_CHAINS, random_seed=i, idata_kwargs={"log_likelihood": True},
        progressbar=False
    )

    ppi = kpt.ProjectionPredictive(model, idata)

    ppi.project()
    cmp_df = ppi.compare()
    result.append(all(cmp_df.index == ["reference", "x", "y", "Intercept"]))
    print("Iteration", i, ":", cmp_df.index)