Performance optimization related to priority queue #76
Description
We use a priority queue for the search stage (growing all the clusters simultaneously), by learning how Sparse Blossom (PyMatching v2) achieves a faster speed than fusion blossom.
Recently I took a closer look at PyMatching's implementation (and got amazed how well their code is structured, which I hope to learn in the future). It turns out that for each vertex, they only add the most recent event from incident edges to the queue. Our current implementation, instead, adds all the future events of all incident edges to the queue. The former adds way less events (>10x) to the queue while still achieving the same effect, so we should definitely learn that.
Credit to Chen Zhao who helped me spot this inefficiency.