Improve seed point reconstruction in seafloor gridding

[jcannon-gplates]

Improve seed point reconstruction in seafloor gridding

This new algorithm naturally deactivates seed points at convergent plate boundaries
without having to compare velocity deltas and distances to boundaries
with empirical thresholds (like the subduction collision parameters).

This comment was written with Claude's help. However, the code is all-human (that'll soon change though).

Summary

• Rewrite collision detection for seafloor seed points — the previous approach
  deactivated a seed point when it transitioned between plates and the velocity delta
  exceeded a threshold (and the point was close enough to the boundary). The new approach
  takes the resolved plate containing a seed point at the current time, resolves that same
  plate at the next time step, and deactivates the point if its reconstructed position falls
  outside the next-time boundary. Any converging boundary will naturally consume seed points
  without requiring velocity thresholds, distance thresholds, or subduction-zone labels.
  This removes the _DefaultCollision, _ContinentCollision, _ReconstructByTopologiesImpl,
  and _ReconstructByTopologicalModelImpl classes entirely.

• Prefer resolved deforming networks over rigid plates when both cover a point — the
  topology lookup is now split into two passes (networks then boundaries) using a shared
  spatial tree, with the network result taking precedence. This correctly handles deforming
  networks that overlay rigid plates.

• Accept PlateReconstruction instead of raw rotation model + topology features —
  ReconstructByTopologies.__init__ and reconstruct_points_by_topologies() now take a
  single plate_reconstruction argument. Topological snapshots are obtained from
  PlateReconstruction.topological_snapshot(), which caches resolved topologies and avoids
  redundant resolving across concurrent workers.

• Deprecate subduction_collision_parameters in SeafloorGrid — the parameter is no
  longer used (collision detection no longer depends on velocity thresholds). Passing it now
  emits a DeprecationWarning; the parameter is removed from __init__ and the class
  docstring is updated to remove references to the old velocity-delta mechanism.

What happens when a topology at the current time cannot be resolved at the next time

This results in inconsistent next-time topology boundaries, which are handled by cloning and extending their sub-segment valid times.

To deactivate seed points at converging boundaries, the code resolves the same plate topology at the next time step and checks whether each seed point's reconstructed position still falls inside it. This requires that all boundary sub-segment features are valid at the next time, but a topology's valid time period often does not include the next time (e.g., at the last time step before a topology disappears). _NextTopologicalFeatures handles this by cloning any boundary section feature (and any topological-line sub-section feature) whose valid time does not include the next time, extending the clone's valid time to cover it (distant past → 0 Ma), and using the clone when resolving at the next time. If the resulting next topology is still inconsistent with the current one — detected by comparing boundary sub-segment counts and checking for unexpected multiple intersections between neighbouring sections, which can distort the resolved boundary shape — _NextTopologicalBoundary falls back to reconstructing the inconsistent sub-segments directly: each current sub-segment geometry is reverse-reconstructed to present day, its valid time is extended to cover the next time if needed, and it is then reconstructed forward to the next time, with the resulting geometries stitched into a replacement boundary polygon.

[jcannon-gplates]

Because this PR relies on functionality that will be in the next pyGPlates release (1.1), I won't merge it until that's available on conda (because the conda version of pyGPlates is used for testing GPlately).