Mixed single-column and multi-column clustering restrictions are not merged correctly

During working on CEP-29, I found that single-column and multi-column clustering restrictions are not merged correctly and cause query failures or incorrect results.

The algorithm used in RestrictionSet.mergeRestrictions is incorrect, as its results depend on the (undefined) order of restrictions in a hashset.
In more complex scenarios, when there are more restrictions to merge, the algorithm results can even differ between the test runs.