fix(device): read the subsystem dict, not its keys

[BombermaG]

subsystem_get() returns a single subsystem dict, but the probe passed it to list.extend(),
which appends the dict's keys instead of the dict.
Listener and namespace were therefore always reported missing, even on a perfectly healthy device.

device_repair() calls this probe twice: once to decide what to rebuild, and once
afterwards to verify. Both calls got the same wrong answer, so:

• the early return that skips an intact stack without spending an attempt (if not missing) could never be reached;
• _def_create_device_stack() is additive and correctly rebuilt nothing;
• the verification probe then reported the same phantom gaps, so a successful no-op was recorded as a failed attempt.

Five of those and the device is marked retries_exhausted, which disables automatic repair for it permanently - including for a real failure later.

Self-repair therefore never confirmed success on any device since it was introduced in 6872469.

R26.3 only. main does not have the self-repair feature yet, so it does not have this bug.

[mxsrc]

Nice catch! With this in mind we can simplify further:
The whole _get_subsys is not required anymore, and subsys can simply be assigned the result of subsystem_get(...). It raises on errors, so the _safe wrapper is not required anymore. The list type is a left-over of the previous filtered subsystem_list.