fix: stale KV fallback + never cache null on /perps cohort fetch

[Flotapponnier]

Summary

• When the worker hasn't written a fresh KV snapshot in < 2 min AND Prometheus is unreachable (always the case on Vercel — VPS Prom not public), fetchPerpCohortRaw() was returning null
• That null was being cached by unstable_cache for 60 s → a single worker hiccup blanked the page for all users for up to a minute
• Same issue for fetchPerpByAssetMatrixRaw (cached empty array for 120 s)

Fixes:

1. Stale KV fallback: if both the fresh snapshot check and the Prom fallback fail, serve the KV snapshot up to 10 min old rather than returning null. Covers brief worker gaps without hiding sustained outages (> 10 min still shows the banner).
2. Never cache null/empty: throw on null/empty so unstable_cache does not store the failure — the next request always retries instead of serving a cached failure for the full revalidate window.

Test plan

• Visit /perps on staging — should show data consistently
• Confirm "temporarily unavailable" banner only appears when worker has been down > 10 min

🤖 Generated with Claude Code