connectd: add --dev-max-wake-delay-ms

[Amperstrand]

The wake-delay watchdog (write_to_subd's empty-queue check) measures the gap between handing a peer message to the subdaemon queue and the subdaemon consuming it, which includes lightningd-master's time to spawn the subdaemon: under CI load that alone can exceed the hardcoded 5000 msec and fail runs with wake delay for WIRE_OPEN_CHANNEL BROKEN (#9268) — tests/test_connection.py::test_funding_cancel_race spins up 100 nodes on a fast runner and hit a 11068 msec window. dev_lightningd_is_slow exists for the same class but is only settable by the memleak dev-RPC.

This makes the threshold a developer option (--dev-max-wake-delay-ms, default unchanged at 5000) so load-heavy tests can raise it; wired through connectd_init. The regression test freezes the accepter's master past the default threshold and expects a clean run with the option raised.

deterministic repro results (master build, regtest pyln) — expand

run1/C2-repro REPRODUCED (pytest failed on broken log, as in CI)
run2/C2-repro REPRODUCED      run3/C2-repro REPRODUCED
run1..3/C1-control PASS (identical flow, no SIGSTOP)
9268/option-b PASS, 9268/in-PR-test PASS (6/6 cells green)

Fuller repro record — the trigger recipe with its two ordering traps, all cell results, and the control arm: https://gist.github.com/Amperstrand/b50c4d604f88016047e104dc65f866c6

Fixes #9268