GraphType cached derived fields + plain-Python graph-op hot paths

[yinoneliraz]

• ComputationEdge: cache computed_hash (edges are hashed on every frozenset build/lookup; fields are immutable).
• GraphType: lazily cached derived fields — nodes, node_to_incoming_edges, node_to_forward_neighbors — plus seed_nodes() so transformations that know the node-set delta (merge_graphs, replace_nodes, replace_source) propagate the cache in O(delta) instead of forcing an O(E) rescan.
• make_computation_node: memoize by callable, skipping repeated inspect.signature work.
• replace_source / replace_destination: single-pass plain loops instead of 4 gamla-combinator passes over the edges.
• merge_graphs, make_first/make_and/make_or, _make_compose_inner, _infer_composition_edges: plain-Python hot paths (kills per-edge curried closures and hidden iscoroutinefunction checks); unbound signature computed once per candidate node using the cached incoming-edges field.
• duplication: use cached nodes/forward-neighbors fields.

Benchmark (bot-building-shaped workload, 4498 edges, 200 replaces): 3141ms -> 677ms total; replace_source 2534ms -> 289ms. All 102 tests pass; canonical graph shapes identical to master.