Read the process table with ps in the box

[splitbrain-bot]

Part of #109 (step 1).

Changes

• docker top is removed. The background probe, the tunnel reconciler and the box-level stop now read the process
  table with one ps -eo pid,ppid,etimes,args in the box (containerProcesses in docker.ts). Before, the probe and
  the reconciler got host pids from docker top, and the stop read the table a second time with ps in the box,
  because a kill in the box needs the box's own pids. containerProcessesFromInside is removed.
• The fallback to a ps format without etimes is removed. It was there for host ps versions that do not know it.
  The box image has procps, so elapsedSeconds is always a number, in ContainerProcess and in the API type
  BoxWork. The dashboard no longer shows a dash for a missing age.
• The probe finds the init by pid 1, and the hold by the sleep infinity with parent pid 1. Before, it found both by
  their command anywhere in the tree, because host pids do not start at 1. So a tini or a sleep infinity that the
  agent started was not counted as work. Now it is.
• doc/threads.md says that the table comes from ps in the box.

Tests

• Orchestrator: 881 passed. The docker top tests are removed. The ps parser test has the age column. The upstream
  test has one fake process table instead of two. A new test checks that a look-alike tini and sleep infinity
  under the agent are work.
• Dashboard: npm run check and the unit tests pass. The e2e tests did not run.
• The ps format and the parser agree with the output of a real box (procps-ng 4.0.4, tini, sleep infinity).

Not tested against a real Docker daemon.

Notes for the next steps

• Step 4, the file service. The probe counts each process in the box as work, except the init, its
  sleep infinity, the adapters, their direct children, the residentProcesses of a harness and the ps of the
  reading. A connected boxes-files is none of these. Without a change, the idle reaper never stops a box with a
  connected file service, and the box-level stop kills the file service. Step 4 must add a rule for it in
  gateway/background.ts, for example a fixed command pattern as READING_ITSELF is for ps. When the disk usage walk
  starts a child process, for example du, the rule must also cover the children of the file service. This was the
  same with docker top.
• Step 5 and step 6, setup at the start. boxes-setup, the agent set install and the home seed run in the box
  before it is ready. While they run, the probe reads the box as busy. The entrypoint itself counts as work until it
  replaces itself with sleep infinity. This lasts some seconds and has no effect on the idle reaper, which acts only
  after IDLE_STOP_MINUTES.
• Step 7, the Runtime interface. The process list is now a plain command in the box through execInContainer.
  It needs no runtime method of its own and becomes run(boxId, ['ps', '-eo', 'pid,ppid,etimes,args']). The call has
  no timeout today. The timeout of run will apply to it.
• Step 8, conformance tests. The process list needs ps from procps with etimes in the box image. A conformance
  test can check that this command works in a new box.
• Step 9, Kubernetes. The pid 1 rule needs the init of the box to be pid 1 in the box. In a pod, this is true while
  shareProcessNamespace is off, which is the default. With it on, pid 1 is the pause process, and each box reads as
  busy at all times. The box pod must not set it. A process started by an exec has parent pid 0 on both runtimes. The
  adapter rules do not depend on this.