DEVX-155: fix: start local dockerd instead of using low-space inner dockerd

When no socket has sufficient space, don't fall back to the low-space
inner dockerd (which will fail on image pulls). Instead, kill the inner
dockerd, clean up its data root to free space, and start a local
dockerd on /dev/shm with vfs storage driver.

Also adds pkill of the inner dockerd and cleanup of its data root
(overlay2, image, volumes, containers) before starting the local
dockerd, to free up the 2.4GB used by the inner dockerd's data.

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>

src/devx/molecule/start_docker.py

@@ -212,15 +212,33 @@ def start_docker_daemon(timeout: int = DEFAULT_TIMEOUT) -> bool:
             return True
         click.echo(f"  Insufficient space ({free_gb:.1f} GB), trying next...")
 
-    # If we found a working socket but with low space, use the last
-    # one as a fallback (better than nothing).
-    for sock in reversed(candidates):
-        if _try_socket(sock):
-            click.echo(f"Using low-space fallback: {sock}")
-            return True
-
+    # No socket with sufficient space found.
+    # Don't fall back to the low-space inner dockerd — it will fail
+    # on image pulls.  Instead, start a local dockerd on /dev/shm.
     click.echo(_("Host Docker not available, starting local dockerd..."))
 
+    # Kill the inner dockerd (started by the CI image) to free its
+    # data root and socket.  The inner dockerd uses the container's
+    # overlay (38G, often 100% full).  Killing it frees up the
+    # socket and any space used by its containers/volumes.
+    with contextlib.suppress(Exception):
+        subprocess.run(  # nosec B603 B607
+            ["pkill", "-f", "dockerd.*--host fd://"],
+            check=False,
+            timeout=5,
+        )
+    time.sleep(2)
+
+    # Clean up the inner dockerd's data root to free space
+    inner_data_root = "/home/grm-ci-runner-*/.local/share/docker"
+    rm_paths = " ".join(f"{inner_data_root}/{d}" for d in ("overlay2", "image", "volumes", "containers"))
+    with contextlib.suppress(Exception):
+        subprocess.run(  # nosec B603 B607
+            ["sh", "-c", f"rm -rf {rm_paths}"],
+            check=False,
+            timeout=30,
+        )
+
     # The existing dockerd (started by the CI image) uses the default data
     # root on the CI container's overlay (38G, often 100% full). We can't
     # kill it (different PID namespace) or bind to the same socket. Use a