Bound blocked dials in the test suite, and add an offline opt-out

[Ticed]

Summary

• tests/inferno/git_integration_test.sh has an environmental skip-guard (INFR-312) that only fires when the dial fails. A connect held open — a host firewall in ask mode, a blackholed route — never fails, so the guard never runs and the emulator sits with no output holding its ports.
• webclient's tlsdial and request() now run dial plus TLS handshake in a child proc raced against a 10s timer, and kill the child on timeout. The kill is load-bearing: a proc blocked in connect() keeps the emulator alive after init() returns — 20s+ past test completion without it, 3.0s with it.
• tcp_test.b's dialTimeout already bounded a blocked connect correctly; it got the same kill so it stops leaking the dialer.
• Adds INFERNODE_TESTS_OFFLINE=1, so a runner declares its environment instead of the suite inferring it from a dial that may never return. run-tests.sh passes -o, the runner writes /env/INFERNODE_TESTS_OFFLINE, and the network tests skip before dialing.

No module interface changed.

Test plan

• Reproduced first, with an accept-and-never-respond listener: discovering refs... then nothing, killed at 45s with no PASS/FAIL/SKIP. A refused dial was run as contrast — the guard fires instantly, isolating the defect to the blocked connect
• After: connect tcp!127.0.0.1!18443: no response within 10000ms, then SKIP: clone failed, at ~10s
• Network present, which is the check that matters — a timeout must not turn a working test into a permanent skip: real clone of octocat/Hello-World passed all 10 steps; webclient_test passed all 4 network tests against real TLS; tcp_test passed Loopback/TcpDialIp/TcpWrite
• /env/INFERNODE_TESTS_OFFLINE verified writable and readable inside the emulator

Checklist

• Commits signed
• No behaviour change when the network is reachable

One thing found on the way and not fixed here: git_integration_test.sh starts ndb/cs, and a running cs keeps the emulator alive after the last test reports — a second, independent reason a suite run hangs at exit. Happy to send that separately; it needs a decision about who reaps it.

[Ticed]

Rebased onto current master (df34b02). Force-pushed cecbdf8 -> df8d227.

git range-diff reports every commit unchanged — same content, same messages,
only the base moved. The five commits that landed underneath are #583, #582,
#566, #578 and #580; none of them touches this branch's files.

Re-verified on macOS after the rebase.

runner.dis -o sets INFERNODE_TESTS_OFFLINE and the networked cases report
SKIP against it, while the loopback case still passes. A dial to TEST-NET
192.0.2.1 now returns bounded instead of hanging: 3004 ms for the tcp case and
10004 ms for the TLS one, "no response within 10000ms". That is the case the
change exists for — a dial that is silently dropped rather than refused used to
hang the suite indefinitely.

#578 fixed the ClusterFuzzLite link break, so Fuzz should now be green here
rather than red for a reason that was never this branch's.