bug: Multicast with IgnoreErrors() leaks router pending entries

[meling]

The following was reported by Magnus Egeland.

Multicast with IgnoreErrors() registers entries in MessageRouter.pending that are never removed. One-way handlers return nil, nil, so HandleRequest sends no response, and the pending entry is never cleaned up. The map grows without bound.

Impact: At 1000 req/s with 4 nodes over 10s, pending reaches 20,000+ entries/node. Unbounded memory growth; any requeuePendingMsgs() call becomes catastrophically expensive.

func TestMulticastIgnoreErrorsLeaksRouterPending(t *testing.T) {
    systems := gorums.TestSystems(t, 3)
    for _, sys := range systems {
        sys.RegisterService(nil, func(srv *gorums.Server) {
            srv.RegisterHandler(mock.TestMethod, func(_ gorums.ServerCtx, _ *gorums.Message) (*gorums.Message, error) {
                return nil, nil
            })
        })
    }
    gorums.WaitForConfigCondition(t, systems[0].Config, func(cfg gorums.Configuration) bool {
        return cfg.Size() == 3
    })
    cfg := systems[0].OutboundConfig()
    for i := range 1000 {
        ctx := gorums.TestContext(t, 5*time.Second)
        gorums.Multicast(cfg.Context(ctx), pb.String(fmt.Sprintf("mc-%d", i)),
            mock.TestMethod, gorums.IgnoreErrors())
    }
    time.Sleep(500 * time.Millisecond)
    for _, node := range cfg.Nodes() {
        if pc := node.PendingCount(); pc > 0 {
            t.Errorf("node %d: pending = %d; expected 0", node.ID(), pc)
        }
    }
}

node 1: pending = 0     ← self-node (local dispatch)
node 2: pending = 1000  ← LEAK
node 3: pending = 1000  ← LEAK
--- FAIL

There exists an easy fix and a better fix. I will soon post the "better" fix, even though it touches more parts of the code.