[tokibito]

Documentation only, following #58. The Running several workers section had a short list of what N costs; this replaces it with a Trade-offs subsection that says what `--workers N` buys and what it costs, so N can be picked deliberately.

• Isolation: one crashed, leaking or hung task costs one worker, and nothing is asked of the task code.
• Memory: N full Django processes with no copy-on-write sharing, since the children are spawned rather than forked; `--max-tasks` as the leak bound.
• CPU: one interpreter and one GIL per worker, so CPU-bound tasks scale across cores and more than the core count only helps I/O-bound ones.
• Polling load and connections, as before, with the LISTEN/NOTIFY broadcast note.
• Stability of the supervisor itself: what happens to the workers if it is killed outright (they are in their own process group and keep running), why that is a non-issue under systemd's default `KillMode` and in a container, and that `requeue_stale_database_tasks` covers the rows either way.
• What a process manager still does better, and that running `--workers N` under one is a reasonable middle.
• Threads inside a worker are a later step tracked in Built-in parallel workers: a supervisor with --workers N, and --threads M inside a worker #25, and on this backend would save memory but not connections.

The CHANGELOG entry for `--workers` under Unreleased gains a sentence pointing at the subsection. No code changes.

Refs #25