Path ids are now located during inference instead of when parsing the alignment paths. With this change all ids is not stored in memory at the same time anymore thus decreasing overall memory usage significantly.
Inference path clusters are now by default inferred from the paths and not the reads. The read based clustering is only needed if multi-maps are used, which is currently not supported. Furthermore, the clustering is now also multi-threaded.
Probabilities are now collapsed during matrix construction. This reduces peak memory for really large clusters.
The following PR contains these major changes: