power: coalesce backlight helper writes, stop blocking on every step

[rumours86]

The bug. Holding a brightness key makes brightness keep changing for seconds after the key is released, and pressing the opposite key has no effect until the backlog drains. Single presses behave fine.

Why. Every Screen.StepUp/StepDown runs up to three synchronous g_spawn_command_line_sync() round-trips — csd-backlight-helper --get-brightness, --get-max-brightness, then pkexec csd-backlight-helper --set-brightness with a polkit round-trip — blocking the daemon's main loop for 60–400 ms per step. Keyboard autorepeat produces a step every ~33 ms, so the calls pile up in the daemon's D-Bus queue and are processed strictly FIFO: brightness "runs on" after release, and an opposite press waits behind the whole accumulated tail.

The fix (same approach gnome-settings-daemon took in gsd-backlight):

• keep the last requested value (backlight_helper_target) and step from it instead of re-reading hardware; answer the D-Bus call immediately;
• write asynchronously (g_spawn_async + child watch) with a one-slot queue: at most one helper in flight, and on completion only the latest requested value is written — a burst of N steps costs 1–2 helper invocations instead of N;
• cache get-max-brightness (never changes at runtime, was re-read via the helper on every step);
• GetPercentage reports the target while a write is in flight, keeping OSD/applet sliders in sync with what the user asked for;
• fix backlight_set_percentage falling through to the helper even after a successful xrandr write (step functions already had the corresponding goto).

Measured on Mint 22.3 (intel_backlight, max=120000), built at tag 6.6.4 and run live:

	before	after
one StepDown D-Bus call	60–90 ms	~1 ms (+ client overhead)
burst of 15 StepDown	~1.4 s serialized pkexec spawns, queue drains long after	all answered in ~15 ms total, exactly 2 helper writes (first + coalesced final)
opposite step right after a burst	waits for the whole tail	takes effect immediately

Helper-write log from the live test showing the coalescing:

backlight write started: 31800
backlight write done, pending=2400   <- 14 steps arrived meanwhile, folded into one
backlight write started: 2400
backlight write done, pending=-1

(the g_message debug lines are not part of this PR)

🤖 Generated with Claude Code

[rumours86]

Following the same standard requested on the cinnamon PRs: now verified against git master as well. Built master csd-power (with master schemas) and ran it live in place of the system daemon on Mint 22.3 / intel_backlight:

• a burst of 15 StepUp calls: every call answered in 14–23 ms (that's mostly gdbus client overhead), final brightness landed exactly at start + 15 × step — no steps lost while stepping from the target;
• the whole burst plus one opposite StepDown produced 3 helper spawns total instead of 16 (first write + coalesced tail + the down step);
• the opposite StepDown right after the burst answered in 12 ms and took effect immediately.

Previously posted numbers were from the same patch built at tag 6.6.4; behavior is identical on master.

🤖 Addressed by Claude Code