fix(extract): finalize detached runs that extract artifacts

[enixCode]

Problem

A detached run with extract never finalized. extractFromVolume's seeder container used AutoRemove and waited on the hijacked attach stream closing, which a volume-bound container does not reliably emit on Docker Desktop / Windows. The run hung forever in running, its container and volume were never cleaned up, and any caller polling the run state (light-run, and through it light-process) looped indefinitely. Every light-process node run hung.

Root cause

The extract seeder relied on the hijacked attach stream emitting end/close to know tar had finished. Volume-bound AutoRemove containers do not reliably emit those events on some Docker builds (the write path already documents this and moved to putArchive; the read/extract path still relied on stream close).

Fix

• seeder: AutoRemove off; wait for the seeder to exit via inspect() polling (awaitSeederExit, the same race-free approach the detached run path already uses), drain the final bytes on an idle timeout (drainIdle) instead of on stream close, and remove the container explicitly (removeSeeder).
• extract: runAlpine and streamTarOut use the new lifecycle with a finally that always removes the seeder.
• listStates: skip sibling files in the state dir that have no id (e.g. cache-usage.json), which made listing runs throw path.join(undefined) (HTTP 500 on GET /runs).

Verification

• 29 targeted tests pass: state unit, volume/extract e2e (incl. a 5 MB file), detached e2e.
• The actual fixed DockerRunner.run detached+extract returns { success: true, extracted: [{ status: 'ok', bytes: ... }] }.

build with cc