Feature: Serialize contraction hierarchy (save/load CH)

[jamescollinharky]

---

Problem

Pandana builds a contraction hierarchy (CH) in the C++ backend when you create a Network. Only the graph is persisted: save_hdf5() saves nodes and edges, but not the CH. So every from_hdf5() rebuilds the contraction hierarchy from scratch, which can take a long time on large networks.

Proposed solution

Add the ability to save and load the precomputed CH so that reloading a network can skip the contraction step.

New/updated API:

• network.save_ch(base_path) — write the CH to binary files (base_path_0.bin, base_path_1.bin, … for each impedance).
• Network.from_hdf5(filename, ch_path=base_path) — load the graph from HDF5 and, when ch_path is given, load the CH from those files instead of running contraction.

Example:

# After building a network once
network.save_hdf5("graph.h5")
network.save_ch("ch")

# Later: load without re-contracting
network = pandana.Network.from_hdf5("graph.h5", ch_path="ch")

Optional extension: make Network picklable (__getstate__ / __setstate__) so the full network (graph + CH) can be serialized with pickle.dump / pickle.load.

Implementation notes

• CH lives in the C++ layer (ContractionHierarchies in libch). The change adds SaveContractedEdges(path) and LoadPreprocessing(path) there, plus Python/Cython wiring and the save_ch / from_hdf5(ch_path=) API.
• Binary format: header (node count, edge count) plus one record per contracted edge (source, target, distance, direction flags, etc.). No version field in the format yet.

I have a working prototype (patch + tests + design doc) and can open a PR against dev if this direction fits the project. Happy to adjust the API or scope based on maintainer feedback.

---