[CI] test-migration: install+createdb+restore in one step (survive per-step-container runners)

[dnplkndll]

What changed

Both migration-test workflows (test-migration.yml baseline + test-migration-enriched.yml) previously split DB provisioning across separate steps: Install postgresql-client → Raise lock limit → DB Creation (createdb) → DB Restore. On a subset of the "oca forks" self-hosted runners each run: step gets a fresh container, so the client installed in the first step is gone by createdb → createdb: command not found (exit 127) and the whole migration test fails for reasons unrelated to the code under test.

This folds install → lock-limit → createdb → restore into a single step per workflow, so the client is guaranteed present where it's used regardless of runner step-container behavior.

• Enriched: the PG16 client path now goes on PATH via in-step export instead of $GITHUB_PATH (which only affects later steps, useless inside one step) — pg_restore still reads the PG16 dump.
• No behavioral change on runners that already persist the filesystem across steps; the cloud_storage_google skip and lock-limit raise are unchanged.

Scope

This hardens the authoritative gate — the ledoent / aggregated pushes and the enriched run. The per-PR test-migration (PR → base 19.0) executes OCA's stock workflow, which has no client-install step at all; that flake is OCA-side and not addressed here. Per-PR tests + pre-commit remain the reliable per-PR signals.

YAML validated; merged-step parses and contains no leftover split steps.