fix(chart): drain s3proxy pods on scale-in instead of killing in-flight uploads

[ServerSideHannes]

Problem

The KEDA connection-demand trigger (argocd-system #807, live since 2026-07-19 ~14:30 UTC) fixed scale-in decisions — the fleet no longer reads idle while busy. But the 2026-07-19 backup run still died at 23:49 UTC with InvalidPart, because scale-in execution is still a hard kill:

• Top-of-hour cron bursts (clickhouse/postgres backups) spike sessions to ~97 → KEDA correctly scales to 30.
• Burst ends → HPA scales 30→6 within ~15 min while Scylla still holds 20–40 long-lived upload sessions.
• Victims get SIGTERM + 30s grace, no preStop. A large UploadPart/UploadPartCopy takes longer than 30s → killed mid-request → orphaned multipart state → InvalidPart on CompleteMultipartUpload.

Fix

Let a terminating pod finish what it's serving:

• preStop sleep 10 — pod keeps accepting while it drops out of the headless-service DNS, so HAProxy (server-template + resolvers) stops routing new connections to it before uvicorn stops listening. Existing connections are untouched by HAProxy when a resolved server goes away.
• terminationGracePeriodSeconds: 600 (was 30, now a values knob) — uvicorn on SIGTERM stops accepting and waits for in-flight requests, then exits immediately. The grace period only needs to cover the longest single request; 600s bounds a slow multi-minute part copy. Pods exit as soon as they're idle — the ceiling costs nothing in the common case.

Chart-only change, no image change needed.