Queries prepared with `OPA.Engine.prepareForEvaluation` do not work when engine has multiple bundles

[dev-rice]

In #11, I found that using queries from an OPA.Engine that had multiple bundles returns unexpected results

@patrick-east narrowed it down to how the store is built up. Data from all bundles is written directly under the data tree, with each bundle overwriting any data that the previous one wrote into the store. When using for ... in directly with a dictionary, iteration order is not consistent, which results in unpredictable query evaluation (as shown by the flaky test in #11).