nx_task_graph.nx

nx_task_graph.nx -- DAG of tasks with explicit dependency edges. Built on nx_thread_pool: each node is a (fn_ptr, ctx) task plus dependency edges to/from other nodes. Scheduling is topo-sort via atomic counters: when a task finishes, it atomic-decrements each successor's `predecessors_done` counter; any successor whose counter reaches `n_predecessors` is now ready and gets submitted to the pool. Initial frontier (nodes with zero predecessors) is submitted by `nx_graph_run`. Why this matters for SSS-class workloads: * ML inference: layer N can't run before layer N-1 finishes. Task graph naturally expresses this without manual barriers. * Render pipelines: vertex -> raster -> pixel -> tonemap stages with parallelism inside each stage. * Build systems: target depends on N source files; ninja is basically a task-graph engine. * MapReduce: map tasks -> shuffle -> reduce tasks. MVP shape -- 8 successors per node ceiling (fixed-size inline), no cycle detection (caller's responsibility), no priorities. All three are next-evolution concerns. Composes against: [[nx_thread_pool_shared_queue]] (executor), [[atomic_intrinsics_real_amo]] (predecessor counters), [[fn_ptr_indirect_call]] (typed task dispatch), [[vyukov_mpmc_channel]] (pool's task queue).