Preload and share Drive map caches across worker processes

[vcharraut]

What

• Add preload_map_cache configuration for PufferDrive.
• Load configured maps once in the parent process before workers start.
• Share immutable road geometry, grid maps, lane graphs, and neighbor caches with workers through fork copy-on-write.
• Add C bindings to preload and release map-cache entries.
• Add lightweight config_only environment initialization to avoid constructing an unnecessary driver environment.
• Correctly manage cache ownership and cleanup across parent and worker processes.
• Use a local fork context when preloading, including when Python 3.14 defaults to forkserver.
• Add C and Python regression coverage for inherited cache reuse and forkserver compatibility.

Why

Each multiprocessing worker previously loaded and retained its own copy of every map. Large multi-worker training jobs therefore duplicated static map data and consumed significantly more RAM.

Preloading static geometry before forking lets workers inherit the same memory pages through copy-on-write. This restores the intended memory reduction without globally changing the application’s multiprocessing start method.

Notes

• Disabled by default with preload_map_cache: 0.
• Requires use_map_cache: 1.
• Non-preloaded environments retain the platform-default multiprocessing behavior.
• Compatible with Python 3.14 and previous Python versions on platforms supporting fork.

[Victorbares]

LGTM.
Nitpick for later : I think we may want to remove use_map_cache if the preload is sufficient