Clustering and consensus

Reads are clustered by address and similarity, then each cluster is reduced to a consensus with an indel-aware full template. V9 adds a native polish kernel (vnx_cl_edit_costs, byte-identical to the NumPy reference) and the opt-in candidate E: ClusterConfig(consensus_template="full", c_indel=4, fill="template"), selected in a pre-registered comparison (see benchmark). The library default is unchanged.