BUG FIX: Shut down relaunched cluster workers when the plan is stopped

[Arthur031221]

A user of plan(multisession), or of plan(cluster, interrupts = TRUE), who cancels a future and then starts another one gets a relaunched worker that plan(sequential) never shuts down; its process and connection stay open until garbage collection, which also produces the "closing unused connection" warning. The same happens with plan(cluster) when a worker died by other means and was relaunched.

When a terminated worker is relaunched (in requestNode() and handleInterruptedFuture()), the new node is stored in the backend but not in clusterRegistry. clusterRegistry$stopCluster() then closes the old, already closed, node. This change adds clusterRegistry$replaceNode() and calls it at both relaunch sites, when the backend's workers are the cluster the registry started. A cluster passed in by the user, for example plan(cluster, workers = cl), is not owned by the registry, so nothing changes for it.

Reproduction, before the change (Linux, one worker, the backend object held so garbage collection cannot hide it):

library(future)
plan(multisession, workers = I(1))
f <- future(Sys.sleep(30)); Sys.sleep(1); cancel(f)
f2 <- future(42); value(f2)
pid <- plan("backend")$workers[[1]]$session_info$process$pid
b <- plan("backend")
plan(sequential); Sys.sleep(1)
tools::pskill(pid, 0L)   # TRUE: worker still running
nrow(showConnections())  # 1

After the change the worker is gone and there are 0 connections.

The new test test-cancel-relaunch-cluster.R counts open connections around this sequence. On the parent commit it fails (before = 0, after = 1, then Error: n <= n0 is not TRUE); with the change it passes (before = 0, after = 0). R CMD check --as-cran on the change ran 112 test scripts; 111 pass, and test-multisession-libpaths.R fails the same way on the parent commit in my environment. The existing cancel, cluster,worker-termination, interrupts-from-worker-itself, value-error-cancels-set, cluster-connection-clashes, plan and reset tests are among those that pass.

This also touches the symptom in #830. The loop from #820 (100 iterations of 10 erroring futures on 2 multisession workers, options(warn = 1), counting the warning lines) printed "closing unused connection" between 7 and 9 times per run on the parent commit across five runs, and 0 times in every run with the change. The leak is intermittent, so I only claim the cancel and relaunch case shown above. I did not check whether other causes of that warning remain.