import_replication.test: interrogate the servers when an RPC deadline expires (#574)

[ohohoreilly]

Makes the next occurrence of #574 diagnose itself. Test-only, no engine change.

Why

The fixture's 240s RPC deadline (#546) fired once on aarch64: the master stopped answering NewFastPrivatePov. The throw unwound into teardown, which SIGKILLed every server, so the run kept only log tails. Those can't tell a wedged master from a dead one, or either from a stuck client i/o thread like #537's. The full reasoning is in #574.

What

• TChildServer keeps a registry of live children.
• Before AwaitAnswered throws, it prints, for each child, Describe() (alive, state [S] or killed by signal N) and per-thread wchan/syscall. Then it prints the same for the test process's own threads.
• import_replication.test: interrogate a master that will not die (#564) #565's DumpWedge is split so the per-thread half (DumpThreads) also reads /proc/self. It stays noexcept and best-effort.

Proven to fire

A throwaway branch forced a 0s deadline on ImportCoreVector, and ci.yml was dispatched on it (run 36958768274; branch since deleted). The make test job printed:

RPC [ImportCoreVector] unanswered after 0s (#537) -- interrogating before teardown (#574)
child 35928: alive, state [S]
  tid 35928  state=S  wchan=sigsuspend  syscall=130 0x7fff323fdd10 ...
  ... 56 more master threads ...
this test process (client side, #537):
  ... 3 threads: main, client i/o, client dispatch ...
end ImportReplication; fail

After that, the fixture failed normally and GracefulShutdownUnresponsiveSlave still ran and passed.