docs(scaling-dive): rewrite with cliff-finding + Qodo fixes

Captures everything learned since the first draft:

- The "250-author cliff" was a measurement artefact from per-IP
  commitRateLimiting + colocated harness. Fixed via the
  etherpad-load-test#105 workflow patch. Real ceiling is ~350-400
  authors on a 4-vCPU GitHub runner.

- apply_mean ballooning at the cliff isn't slow code — it's OS
  preemption (7+ cores of work on 4 vCPU). Application-level JS
  rearrangement can't reach it.

- Two changes hold up under the dive: fan-out serialization
  + NEW_CHANGES_BATCH (#7768, 70% p95 drop at 200 authors) and
  historicalAuthorData cache (#7769, neutral on dive but real
  production thundering-herd fix at join time).

- Four directions didn't pan out: WebSocket-only transport, heap
  bump, message-level batching alone (#7766 closed), and
  rebase-loop prefetch (#7770 closed). Each has a one-line cause
  documented for the record.

- Engine.io transport-level packing (#7767) is the meatiest
  untouched lever — sending multiple packets per WebSocket frame
  the way polling already does via encodePayload.

Qodo-flagged corrections incorporated:
1. The new instruments are Histogram + Counter + Gauge, not
   "three counters" — labelled correctly.
2. The lever-3 line reference now points at updatePadClients
   (lines 985-999) where NEW_CHANGES actually emits, not the
   wrong line 627 (handleSaveRevisionMessage).
3. Lever 3's results are written up against measured data, not
   "deferred".

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: JohnMcLear