fix(proxy): filter on-demand wake/sleep by node_id + stop leaking detail in 503 bodies

[dviejokfs]

Two follow-ups from PR #124 (on-demand first-request 503 fix). Closes #126, closes #127.

#126 — multi-node wake/sleep started/stopped ALL containers locally

OnDemandManager::do_wake and sleep_environment loaded every deployment_containers row for the deployment with no node_id filter, then started/stopped each one via the local Docker ContainerLifecycle. On a multi-node cluster, containers whose node_id points at a remote worker don't exist on the local daemon → start_container fails → the partial-wake revert fires → the request gets the wake_failed 503. So scale-to-zero wake only worked correctly on single-node deployments.

Fix:

• OnDemandManager now carries local_node_id: Option<i32>. A container is local iff node_id IS NULL (the deploy pipeline only stamps a node_id for remote workers, so control-plane-local containers carry none) or node_id == local_node_id.
• do_wake partitions containers and starts only the local set. Remote-owned containers are skipped with a warn! (the worker-side wake RPC doesn't exist yet). A fully-remote environment has nothing this proxy can start, so it reverts sleeping=false and returns an error instead of falsely reporting a successful wake.
• sleep_environment is symmetric: it stops only local containers and leaves remote ones for their own node's idle sweep.
• The control plane passes local_node_id = None (it has no self node row; its locally-deployed containers carry node_id = NULL).

Note: remote containers are not woken by this change — the worker-side wake RPC is a separate, larger multi-node feature. This fix makes the local path correct and honest (warn + skip / error) instead of silently breaking every multi-node wake.

#127 — 503 bodies leaked environment_id and internal error strings

The on-demand wake 503 responses in proxy.rs are served to unauthenticated clients (a sleeping env has no auth context yet) and disclosed internal detail.

Fix:

• Dropped environment_id from all three 503 bodies (wake_throttled, wake_pending, wake_failed) — clients key retries off Retry-After, not the id.
• wake_failed no longer interpolates e.to_string() (the OnDemandError Display string can carry container/deployment context). It returns a static message; the detailed error is still logged server-side via error!(... error = %e ...). This also removes the fragile .replace('"', ...) JSON escaping.

Tests

• Reworked the two pre-existing tests that asserted the buggy "start/stop all" behavior (test_wake_multiple_containers_multi_node, test_sleep_stops_multiple_containers) plus the lifecycle test, to assert local-only behavior.
• Added: test_is_local_container_node_filter, test_wake_starts_only_local_containers_skips_remote, test_wake_all_remote_containers_errors_and_reverts, test_sleep_stops_only_local_containers_skips_remote.
• cargo test --lib -p temps-proxy — 311 passed. cargo check --bin temps clean.