Speed up CI service health checks with 1s start-interval probes

[cnkk]

Two changes to cut "Initialize containers" time in Elixir CI, based on profiling real runs (details below).

1. Fine-grained health-check probing. Since #6571 made the service health checks real, the step went from ~10s to a 31-50s median. Postgres is ready in ~1-2s but with `--health-interval 10s` Docker runs its first probe only at t=10s, so the healthy flag can't flip earlier and the runner detects it at ~14s. `--health-start-interval 1s` (Docker Engine 25+; runners are on 28) probes every second during the start period so the flag flips at actual readiness:

```yaml
options: >-
--health-cmd "pg_isready -h 127.0.0.1"
--health-start-period 30s
--health-start-interval 1s
--health-interval 10s
--health-timeout 5s
--health-retries 5
```

Postgres also gets `-h 127.0.0.1` so the probe checks TCP instead of the unix socket, since the official image's temporary init-phase server listens only on the socket. Steady-state cadence (10s interval, 5 retries) is unchanged.

2. Non-alpine ClickHouse image. A 6-variant profiling matrix on `blacksmith-4vcpu-ubuntu-2404` showed the alpine image, served from the Docker container cache, pays ~29s of lazy layer hydration on first use per VM (~10s before the entrypoint even executes), while disabling system log tables and tmpfs on `/var/lib/clickhouse` had zero effect on boot. The non-alpine `clickhouse/clickhouse-server:25.11.5.8` bypasses the cache (17s registry pull, overlapped with the postgres wait) and starts in ~0.6s, netting roughly 12s per service-backed job. Caveat: if the container cache later picks this image up, the hydration cost may return; the durable fix is platform-side.