Add allocation cache for ConvexConvexSutherlandHodgman clipping

[asinghvi17]

Summary

ConvexConvexSutherlandHodgman is the fast path for convex-convex polygon intersection, the kind of routine that gets called in a hot loop (e.g. coverage/regridding over many cells). Previously every call allocated a fresh vector per clip edge — m+1 vector allocations plus growth reallocations for an m-gon clip polygon. Since the algorithm is purely iterative (one pass per clip edge), all intermediate state can live in reusable buffers.

This PR adds an exported SutherlandHodgmanCache{P} with two ping-pong buffers (shared by both manifolds) plus clip/subject buffers for the spherical path. It is threaded through as a keyword argument:

alg = GO.ConvexConvexSutherlandHodgman()
cache = GO.SutherlandHodgmanCache(alg)
for (a, b) in polygon_pairs
    result = GO.intersection(alg, a, b; cache)
end

• cache = nothing (the default) allocates fresh buffers per call — identical behavior to before, always thread-safe.
• The returned polygon is copied into a fresh, exactly-sized ring, so results never alias the cache and remain valid after the cache is reused.
• A cache with the wrong point type for the manifold/float combination throws a descriptive ArgumentError (constructors SutherlandHodgmanCache(alg, T) / SutherlandHodgmanCache(manifold, T) get this right automatically).
• Thread safety: deliberately left to the caller — one cache per task. This is documented in the docstring; a task-local convenience could be layered on later.

Octagon-octagon benchmark: 928 bytes/call uncached → 448 bytes/call cached (just the returned polygon).

Test plan

• Existing ConvexConvexSutherlandHodgman tests pass unchanged
• New Cache testset: cached vs uncached result equality (planar + spherical), aliasing regression (reusing a cache must not corrupt previously returned polygons), wrong-cache-type ArgumentError, and a post-warm-up allocation bound

🤖 Generated with Claude Code