Add OSM road benchmark (pyogrio / Geofabrik extracts)

[tomalrussell]

Motivation

• Provide a reproducible benchmark for issue Measure performance with ~1M node/edge network #61 that measures snkit topology performance on realistic OSM extracts.
• Avoid fragile pyrosm usage by reading the GDAL OSM lines layer via pyogrio to be robust in different environments.
• Cache downloads atomically under .benchmark-data and exclude them from measured intervals to ensure repeatable timings.
• Produce machine-readable results including platform and Python metadata so benchmarks can be archived and compared.

Description

• Add benchmarks/benchmark_osm_roads.py, which downloads Geofabrik extracts, reads highway ways via geopandas + pyogrio, and times four steps: reading highway ways, creating endpoint nodes, assigning IDs, and assigning topology IDs, and supports all and --output for JSON reports.
• Implement atomic cached downloads with cleanup of partial files to .benchmark-data and add .benchmark-data/ to .gitignore.
• Replace the previous pyrosm-based approach with read_highway_ways that filters GDAL lines for highway IS NOT NULL and sanitises geometries.
• Add benchmarks/README.md documenting how to run the benchmark and add the benchmark optional dependency (pyogrio>=0.7) to pyproject.toml.

Testing

• Verified benchmarks/benchmark_osm_roads.py compiles with python -m py_compile and passes ruff formatting and lint checks.
• Confirmed the GDAL OSM driver is available via pyogrio.list_drivers() on the environment used for testing.
• Exercised the full benchmark pipeline with a synthetic read_highway_ways replacement and validated expected edges/nodes counts and that timing fields read_highway_ways, create_endpoint_nodes, add_ids, and add_topology_ids are present.
• Attempted real downloads but the execution environment blocked outbound HTTP(S) (HTTP 403 via proxy), so live Oxfordshire/France runs could not be completed and no real PBFs were produced; partial-download cleanup was verified.
• Ran the test suite which produced 23 passed, 1 skipped, 1 failed; the single failure is unrelated to the benchmark and is caused by the optional networkx package not being installed in the test environment.

---

Codex Task