WIP: Add clustering functions to C API

[dbaston]

This PR resurrects the C API that was removed from #688 and changes it to return an array of cluster IDs instead of a GeometryCollection, as suggested in #823.

This is clearly a more flexible API than the one implemented in e464b79, but it becomes somewhat awkward to construct geometries representing the clusters (see function construct_clusters in the test suite).

I am a bit averse to adding new types to the C API, but I am wondering if something like the following would be an improvement.

struct GEOSClusterInfo;

// return cluster info or NULL on error
GEOSClusterInfo* GEOSClusterDBSCAN(const GEOSGeometry* g, double eps, unsigned minPoints);

unsigned GEOSClusterInfo_getNumClusters(const GEOSClusterInfo* clusters);
unsigned GEOSClusterInfo_getClusterSize(const GEOSClusterInfo* clusters, unsigned i);

// Return an array with the input indices associated with the cluster i
unsigned* GEOSClusterInfo_getIndicesForCluster(const GEOSClusterInfo* clusters, unsigned i);

// Return an array with the cluster id associated with each input index
// Input indices not associated with a cluster will have a cluster id of GEOS_CLUSTER_NONE
unsigned* GEOSClusterInfo_getClusterForIndices(const GEOSClusterInfo*);

void GEOSClusterInfo_destroy(GEOSClusterInfo*);

[dr-jts]

A suggestion for naming of a couple of the methods, with clarifying comments:

// Return an array with the input indices associated with the cluster i
// Size of array is given by GEOSClusterInfo_getClusterSize
unsigned* GEOSClusterInfo_getClusterElements(const GEOSClusterInfo* clusters, unsigned i);

// Return an array with the cluster id associated with each input index
// Size of array is equal to the number of elements in the input geometry
// Input indices not associated with a cluster have a cluster id of GEOS_CLUSTER_NONE
unsigned* GEOSClusterInfo_getClusterIds(const GEOSClusterInfo* clusters);

[dr-jts]

Rather than adding another type to the API, perhaps a function similar to construct_clusters could be provided?

That has the downside that all clusters have to be materialized at once, rather than individually. But perhaps this is fine for the anticipated use cases?

[pramsey]

I'm OK with either approach, but I think the extra functionality of the added object is pretty good, and it's supporting 4 different functions, so it's not like it's an extra "junk object" or something.

[dbaston]

Thanks for the feedback. I've implemented a new version with the custom type, but I don't want to rush it out before Friday. I'm OK leaving this until 3.14.

[dr-jts]

There is at least one other non-clustering algorithm that could benefit from this approach - GEOSLIneMerge (and GEOSLineMergeDirected). Returning ids of merged linestrings will support creating a PostGIS ST_LineMergeWin window function, which allows maintaining attribution for merged lines.

Given this, is it worth renaming the API so that it's more general than just clustering? The term Group could be used, giving:

struct GEOSGroupInfo;

// return cluster info or NULL on error
GEOSGroupInfo* GEOSClusterDBSCAN(const GEOSGeometry* g, double eps, unsigned minPoints);

unsigned GEOSGroupInfo_getNumGroups(const GEOSGroupInfo* groups);
unsigned GEOSGroupInfo_getGroupSize(const GEOSGroupInfo* groups, unsigned groupId);

// Return an array with the input indices associated with the group groupId
// Size of array is given by GEOSGroupInfo_getGroupSize
unsigned* GEOSGroupInfo_getGroupItems(const GEOSGroupInfo* groups, unsigned groupId);

// Return an array with the group id associated with each input index
// Size of array is equal to the number of elements in the input geometry
// Input indices not associated with a group have a cluster id of GEOS_GROUP_NONE
unsigned* GEOSGroupInfo_getGroupIds(const GEOSGroupInfo* groups);

void GEOSGroupInfo_destroy(GEOSGroupInfo*);

Or, LineMerge can be though of as a kind of clustering, so an alternative design choice is to use this API and name the GEOS C API function appropriately (say GEOSClusterLineMerge).