Support CSR sparse matrices from scipy.sparse

[daxpryce]

The CompactNetwork is ultimately just a CSR sparse matrix. I had no idea at the time, and now I do, and I feel kind of embarassed. There's probably already csr matrices already in crates.io and I didn't even think to look. Ah well, older and wiser now I guess.

Anyway, if we could take in a csr sparse matrix directly, especially if it's zero copy, I suspect leiden's speed improvements will increase dramatically. We waste so much time copying one set of edgelist-with-nodes-as-strings from python's gil into our own version of Edge, which we then use to build our CompactNetwork, and for some people their network was already in CSR matrix form AND had integral nodes AND they were already doing their own bookkeeping AND we didn't need to make this many dang copies.

Anyway.

Success Gates:

• leiden(...) or leiden_csr(...) and modularity(...) callable that uses scipy.sparse.csr_matrix as an input type
• cannot require a specific version of numpy or scipy (ideally we can support multiple major versions of each, with unbounded minor/bugfix versioning). In other words, numpy 1.x and numpy 2.x.
• zero copy of the matrix, while still not running afoul of the gil. unsure if this is possible, but it is the goal.
• zero copy to use the csr matrix as the CompactNetwork in our current leiden/hierarchical_leiden/modularity
• Performance benchmarks including:
  • time for graspologic.leiden to compute a network partitioning from a network that was originally in networkx.Graph() form with string nodes (actual node values should be integer values)
  • time for graspologic_native.leiden to compute a network partitioning from a network that was originally in Vector form with string nodes (actual node values should be integer values)
  • time for grasologic_native.leiden_csr to compute anetwork partitioning from a network that is entirely csr sparse matrix (no strings involved!)

[daxpryce]

also, maybe support it. looking into the scipy impl it doesn't seem nearly as straight forward as the direct numpy support in #57

[daxpryce]

in theory we do something like (pseudocode):

def leiden_csr(sparse: scipy.sparse.csr_matrix, ...) -> npt.NDArray[np.uint64]:
   compact_network_nodes = sparse.indptr  # ndarray of offsets ala CompactNetwork.nodes
   compact_network_neighbors ~= zip(sparse.indices, sparse.data)
   # todo: we need to sum the self links edge weights and note it for the CompactNetwork. I cannot for the life of me remember why. it probably matters for real tho
   # make our compact network out of references to these
   # call leiden in network_partitions crate

Better would be if we use a sprs csr sparse matrix instead of CompactNetwork, and we can write python .py code to handle the graph labels -> number stuff without bothering graspologic-native with it.

Just riffing on some ideas here.