Deterministic merging #5
Description
The way merging is performed is currently nondeterministic - the order in which candidate merges are performed depends on the order of the dataframe which is not always sorted, and can therefore depend on the order in which parallel computations complete.
It is highly desirable that the output be fully deterministic.
To achieve this, we should make sure that the candidate merges are always sorted. The sort order may differ between different merge heuristics (e.g. exact vs locus), but they should always have a total order.