Modernizing Mesa-Geo: PropertyLayer Refactor, Cell/Agent Decoupling & Visualization Overhaul

Mesa-Geo currently suffers from three structural problems that make it impractical at realistic geographic scale: a memory bottleneck from instantiating one Python object per raster pixel, a fractured coordinate system that silently conflates three incompatible reference frames, and a broken Cell/Agent abstraction that floods the agent registry with millions of ephemeral entries. This project resolves all three problems through four focused changes: refactoring RasterLayer onto a PropertyLayer NumPy backend, decoupling Cell from mesa.Agent through a weakref identity cache, delivering a RasterCellCollection aggregation API with scipy.ndimage-powered spatial statistics, and overhauling the visualization pipeline to render rasters as a single ipyleaflet.ImageOverlay instead of millions of polygon draws. All work is aligned with Mesa 4.0's unified reactive architecture, positioning Mesa-Geo as the geospatial pillar of Mesa's next generation rather than a legacy bolt-on. Core deliverables: PropertyLayer-backed RasterLayer, decoupled Cell with three explicit coordinate properties, RasterCellCollection aggregation API, SolaraViz-compatible make_geospace_component, a geo predator-prey benchmark model, and a full Mesa-Geo 3.x migration guide. Stretch goals: H3 hexagonal grid support, a GeoSpace redesign proof-of-concept (#2585), and a Kalman filter temporal state estimation API for PropertyLayer.