Fix CUDA rasterize MAX_ISECT overflow and optimize classify_geometries

- Increase GPU scanline fill MAX_ISECT from 512 to 2048 to handle dense
  polygon inputs (1000+ irregular polygons would silently drop edges,
  causing wrong pixel values and missing coverage)
- Add host-side warning when active edge count exceeds the GPU limit
- Vectorize _classify_geometries using shapely 2.0 array ops instead of
  per-geometry Python calls (178x faster for the classification step)
- Defer bounds computation in _parse_input so callers who supply explicit
  bounds skip the unnecessary bbox scan
- Add cross-library rasterizer benchmark (benchmarks/rasterizer_benchmarks.py)
  covering 10 geometry types with consistency checks and markdown output

Author: brendancol