feat(deploy): autoscale the WS gateway

im HPA: min 2 / max 6, CPU 70% as the primary signal and memory 80% as the
connection-count guard (against the 128Mi request that is ~102Mi, and each
live socket costs its buffers plus its own pubsub pump; the 512Mi limit is
deliberately not the trigger).

Scaling out is free here: cross-pod fan-out already goes through Redis
pub/sub (Realtime::publish_room / subscribe_room), and the per-process
rooms/users maps only hold this pod's own sockets, so a client on pod A
keeps receiving events published by pod B. Scaling in is the one that
costs something - the retired pod's clients must reconnect - so scale-down
is much slower than apihost's: a 600s stabilization window and at most one
pod per 120s, while scale-up reacts within 30s (+100% or +2 pods per 60s).

A PDB (minAvailable 1) now keeps a node drain from taking both gateways at
once. The image tag stays sha-20a153a: these commits only touch chart/ and
deploy/, and the rendered im PodTemplate is byte-identical to the live one,
so the upgrade rolled nothing and no websocket was dropped (verified: same
pod names and ages, 0 restarts, and forcing the Deployment to 1 was
restored to 2 by the HPA within 6s).