QFw: install the service-plane manager's stop handlers before it starts

[DougSO]

Fixes #135.

The foreground service-plane manager wrote manager_pid to its state file and printed its state, and only then installed its SIGTERM and SIGINT handlers. A caller that sees manager_pid takes the manager as up, as test_foreground_run_stops_on_sigterm does when it sends SIGTERM. A signal in between took the default action:

• the manager died with rc -15;
• stop() never ran, so its components were left running;
• its buffered stdout was lost, hence the empty output in the CI failures.

That is the test's intermittent CI failure, 4 of the 15 failed QFw runs since 2026-09-15.

The fix

run() installs the handlers first, and restores them in a finally that covers start() too. The handlers only set the stop flag:

• A signal after the plane is up stops it, as before.
• A signal during start-up is honoured once start() returns, and start()'s own timeouts bound that wait. The plane is then stopped cleanly.

Before, a SIGINT during start-up raised KeyboardInterrupt. That went past start()'s cleanup, which catches Exception only, so the components it had already started were left behind.

Tests

Two new tests wrap a module function so the manager sends SIGTERM to itself as the function returns:

• test_a_sigterm_just_after_ready_still_stops_the_plane wraps _print_state, which is the race behind the CI failures;
• test_a_sigterm_during_start_stops_the_plane_once_it_is_up wraps start.

The window is hit every time, so these are deterministic. Without the fix, both fail 3 of 3 runs with assert -15 == 0, the CI signature. With it, both pass, and so does test_foreground_run_stops_on_sigterm.

In the container, on main f2c8e93: mock 619 passed and 1 skipped, unit 130 passed, tests/qiskit 35 passed, ci-syntax clean. That holds with and without the container's telemetry environment.