feat(agent): send the CPU, memory and disk use of each site, and conform to contract v0.5.0

[nabil1440]

Closes #45.

Part of the monitoring agent stack. This PR is on top of #49. Contract v0.5.0, "sites". After this PR, the agent conforms to contract v0.5.0, and the README pins it.

What changes

• New optional sample field sites: one item for each Docker Compose project in the home folder of the server user.
  • The agent lists the running containers (GET /containers/json), and groups them by the label com.docker.compose.project.working_dir.
  • It keeps a group when the parent of the folder is the home folder. directory is the last part of the folder, for example example.com or .fly.
• CPU: the usage_usec deltas of the cgroup v2 cpu.stat of the containers, as a share of all the CPUs.
  • Only a container that both ticks saw, with the same id and the same cgroup, counts. A container that started or restarted in the minute is left out. docker restart keeps the id, but makes a new cgroup.
  • The time of the minute is the time between the two reads of the cgroups.
• Memory: memory.current minus inactive_file, at the tick. Both the systemd and the cgroupfs cgroup drivers are supported.
• Disk: at most one time each hour, a background walk measures each project folder.
  • It does not follow a symbolic link, and does not go into another file system. It counts a file with more than one hard link one time.
  • It skips what it cannot read, and writes a warning one time for each directory. The later walks log it at debug level. On a FlyWP server, the databases in ~/.fly/database belong to the container user, so .fly always has such files.
  • It runs on a thread with the idle I/O priority.
  • The next sample carries the result one time. The other samples have null.
  • Each folder has a limit of 5 minutes. A walk that hangs does not stop the next walks.
• null and []. sites is null when the agent cannot read Docker, on cgroup v1, or without a home folder. It is [] when Docker runs and no project matches. The queue on disk keeps the two apart.
• The agent keeps at most 1000 items, and drops an item whose directory is empty or longer than 255 characters.
• The pin in the README, the package documentation and wire moves to v0.5.0.
• A short first minute (see feat(agent): read the server each 10 seconds and send the peaks of the minute #46) also reads the containers, so the next sample has the CPU of each site.

Cost

• With many sites, the queue on disk grows: about 100 bytes for each site in each sample. For 100 sites and a full queue of 24 hours (1440 samples), that is about 14 MB.

Tests

• A fake Docker Engine and a fake cgroup tree: the grouping and the home rule, CPU and memory, a container with a new id, a container that restarted with the same id, the cgroupfs driver, the cases of null and [], and / as the home.
• The disk: sent one time, a walk that fails, a walk that hangs, a sample that fails before the result goes, symbolic links, hard links, and a folder that cannot be read.
• make check passes.