fix(desktop): raise default warm backend pool cap from 3 to 8

[kokhlo]

Summary

Desktop's warm-backend pool capped concurrent spawned bot backends at 3 by default. On a multi-bot roster (the reporter runs 6: 1 remote + 5 local) the LRU cap evicts the least-recently-used idle backend as soon as a 4th is opened, so earlier bots pay a cold start every time they're revisited — matching the "ones I created first become no longer accessible until I re-prompt Hermes" report in #102822.

• The pool was already fully configurable (Settings → Advanced "Warm Bot Backends", clamp range 1–64, HERMES_DESKTOP_POOL_MAX env fallback); only the default was stale — it preserved the original hard-coded ceiling from before multi-bot rosters were common.
• Raised the default maxBackends 3 → 8 in both authoritative places: apps/desktop/electron/pool-limits.ts (main process: pool + persisted preference) and its hand-mirrored renderer copy apps/desktop/src/store/pool-limits.ts (Settings rows + prewarmProfileBackend saturation guard). Values already persisted in userData are untouched — users who lowered the cap keep theirs.
• At ~60MB per warm backend the new default budgets ~480MB across 8 bots, in line with machines that run Desktop with a multi-bot roster; single-bot users still spawn exactly one backend — the cap only bounds concurrently warm ones.
• Added a pinning test asserting the default stays a multi-bot-sized pool (>3), with a comment explaining the two mirrored constants must not drift.

Testing

• npx vitest run electron/pool-limits.test.ts --project electron — 9 passed (incl. new pinning test)
• npx vitest run electron/pool-spawn-coordinator.test.ts electron/pool-eviction.test.ts --project electron — 22 passed
• npx vitest run src/store/profile.test.ts --project ui — 20 passed
• npx tsc --build tsconfig.electron.json — exit 0
• npx tsc -p . --noEmit (renderer) — exit 0
• eslint on all three touched files — exit 0

Fixes #102822

[shellybotmoyer]

LGTM. Clean 3-file fix (+16/-4) raising the warm backend pool cap from 3 to 8. Reasonable default for desktop use cases where multiple model providers may be configured. Straightforward config improvement.