Fix scanline row gaps, speed up rasterize, add resolution/like/merge (#991)

* Fix scanline gap bug and optimize rasterize performance

The scanline fill had a row-clamping bug (floor(r1)-1 instead of
ceil(r1)-1) that left every other row empty in curved polygons.

Also adds three performance improvements:
- Vectorized edge extraction via shapely 2.0 array ops (~13x faster)
- Sorted edge table with binary search in scanline kernel (~7-50x faster)
- Single-pass geometry classification to avoid redundant iteration

* Add resolution, like, and merge parameters to rasterize

- resolution: derive width/height from a pixel size (scalar or (dx, dy) tuple)
- like: copy grid dimensions, bounds, and dtype from a template DataArray
- merge: configurable overlap strategy (last/first/max/min/sum/count)
  with per-pixel dispatch via _merge_pixel helpers (CPU and GPU)
- geometry-ID-based scanline ordering so overlap merge follows input order

* Active edge list, written-mask, GeometryCollection, lazy cuda

- Scanline kernel now maintains an active edge list instead of
  scanning all edges up to the binary-search cutoff. For 2000 small
  circles this drops edge checks from 19.8M to 40K (497x reduction).
- Replaced NaN sentinel with a separate int8 written-mask so merge
  modes (first/sum/etc.) work correctly when burning NaN values.
- GeometryCollections are recursively unpacked instead of silently
  dropped.
- GPU kernels are compiled lazily on first use_cuda=True call instead
  of at import time (rasterize.py no longer does `from numba import
  cuda` at module level).

* Vectorize point/line extraction, clip lines to bounds, CSR GPU scanline

Point and line extraction now use shapely 2.0 vectorized ops
(get_parts/get_coordinates) instead of Python loops. Line segments
are clipped to raster bounds via Liang-Barsky before Bresenham,
cutting iteration from ~20K to ~100 steps for far-off-screen lines.

GPU scanline kernel reads active edges from a precomputed CSR
structure instead of binary-searching + scanning dead edges.

* Renumber user guide notebooks to remove duplicates

Three notebooks shared prefix 17 and two shared 25. Spread the
triple-17s across 17-19 and shift everything after by +2. Gave
the unnumbered zonal_crosstab notebook prefix 30.

Author: brendancol