feat(worker): /proc-based external sampler for per-task RSS and CPU

[albanm]

Adds a parent-side /proc-based resource sampler that runs once per task slot, so the
worker keeps reporting per-task RSS and CPU usage even when the child's in-process
df-mem sampler is starved by a CPU-bound plugin.

What changed:

• New worker/src/utils/proc-stat.ts reads VmRSS from /proc/<pid>/status and
  utime/stime from /proc/<pid>/stat; computes a CPU ratio (1.0 = one full core).
• New worker/src/task/external-sampler.ts runs the reader per slot at
  WORKER_TASK_MEMORY_SAMPLE_INTERVAL_MS, started in iter() after child.spawn()
  and stopped in the close/error handlers.
• The external sampler becomes the authoritative writer for the
  df_processings_process_resident_memory_bytes{kind="task"} gauge; the in-process
  df-mem writer suppresses RSS when the sampler is active.
• New gauge df_processings_process_cpu_usage_ratio{kind="task",slot=…}.
• diagnoseExit now takes a lastExt: ExternalSample | null and renders the
  external RSS/CPU lines before the (possibly stale) child sample in oom-host
  diagnostics; the French user message gets the same treatment.
• New config knob WORKER_TASK_EXTERNAL_SAMPLER_ENABLED (default true,
  auto-disabled at boot on non-Linux).
• Tests: unit specs for proc-stat, external-sampler, and the updated
  exit-code; one e2e (memory-oom.e2e.spec.ts) driving a new processing-cpu-leak
  fixture that busy-loops the child while the parent keeps reading /proc.
• Also bundled: an exit-time df-mem debug write + text-warning/
  text-medium-emphasis color fix in the run-logs list (commit 178204b).

Why: CPU-saturated tasks make the in-process df-mem sampler go stale, so the
lastMem attached to an oom-host diagnostic could be many seconds old at the
moment of the OS kill. Parent-side /proc reading sidesteps the child's event
loop entirely.

Regression risks:

• diagnoseExit signature gained lastExt (5th arg). In-tree callers + tests are
  updated; flag for any out-of-tree consumers.
• memReporter.stop() is now async and is awaited in task.ts's finally.
  The exit-time debug write now blocks task tear-down until the mongo log write
  resolves. A wedged mongo will stall task cleanup; consider a short timeout
  around the awaited debug call.
• worker.task.externalSamplerEnabled is required in config/type/schema.json.
  Deployments with an explicit pinned config will need to add the key (default
  is true in default.mjs).
• If the external sampler stops mid-run (e.g. /proc read fails), the in-process
  writer is not re-enabled — the RSS gauge for that slot freezes at the last
  external sample until the next run. Acceptable per the design, but undocumented
  in the "Failure modes" section.
• E2e fixture busy-loops for ~25 s against a 90 s timeout; CPU-ratio threshold
  (> 0.1) may flake on heavily-loaded CI. RSS-fallback assertion mitigates.