fix(services): poll recorder state when utils are open (#1827)

* fix: sync recorder state with recordings started outside the shell

The recorder only checked for a running gpu-screen-recorder proc when the
shell itself started, stopped or paused a recording. A recording toggled by
keybind therefore left the utilities card out of sync: after an external stop
the card kept showing REC with the timer counting up, and an external start
was only noticed when the card tried to start another recording.

Poll pidof once a second while something is actually showing the state. The
card takes a Ref on the service, so nothing is polled unless the utilities
drawer (or sidebar) is open.

Commands from the shell run pidof themselves, and the proc does not
start/stop the instant the command is dispatched - `caelestia record` waits
for the recorder to finalise the file before exiting, and for region
recordings it waits on slurp before spawning it at all. Poll results are
therefore only applied once they agree with the state we set, or after 10s if
they never do, so a proc that is still shutting down is not mistaken for a
new recording and a command that silently failed does not leave the card
stuck.

Also gate the elapsed counter on the running/paused state again - the
Connections doing that was commented out, so the timer kept counting up while
a recording was paused.

* services: track the record command instead of a timeout

Review feedback from @dxnnv on #1827.

`caelestia record` owns the whole transition: it blocks on slurp for region
captures, and when stopping it waits for the recorder to finalise the file
before exiting. Neither has a bound the shell can assume, so the 10s backstop
could expire mid-command and let a poll read a recorder that was still
shutting down as a brand new recording.

Running the command through a Process instead means its lifetime is the
signal. Poll results are ignored while it is running, and reconciled when it
exits, so no timeout is needed.

Also move the `enabled` binding below the function declaration in the
Connections, which the section-order convention requires, and suppress
signal-handler-parameters on the new onExited the same way checkProc already
does - without it qmllint fails on the unresolved QProcess::ExitStatus.