Install matched kernel-devel in the GPU lane and reboot once

What

The GPU lane's driver step installs kernel-devel-matched and the unversioned kernel-headers, and its kernel-update reboot happens at most once.

Why

startup.sh installed kernel-headers-$(uname -r). Fedora publishes no versioned headers package, so the install always failed, the fallback updated the kernel and rebooted, and the next boot failed the same way: a reboot loop with no driver.

NVIDIA's driver installation guide gives the install for Fedora and AlmaLinux 9, Preparation step 2: dnf install kernel-devel-matched kernel-headers (https://docs.nvidia.com/datacenter/tesla/driver-installation-guide/fedora.html, almalinux.html). kernel-devel-matched pulls the devel package for the running kernel; the headers package is not tied to a kernel version.

How

• gce/startup.sh: the install line above. When it fails because the repository dropped the running kernel, the script installs the current kernel, writes /var/lib/sim-gpu-kernel-updated and reboots. A second failure finds the marker, signals failed and exits, so up reports it and nothing loops.
• gce/operations.md: the reboot sentence now says when it happens. Word budget stays at its ceiling.

No diagram: this is one install line and one guard, with no new flow.

Boundaries

Untested on a fresh rig: no live GPU create ran for this change. bash -n, shellcheck and just ci pass.