worker: honour backend_factory when reporting the machine's backend

[pradipta-lyzr]

What

run_worker resolved the machine's backend name via select_backend("auto") at
registration time, ignoring an explicitly supplied backend_factory.

-be_name = select_backend("auto").name
+be_name = (backend_factory() if backend_factory else select_backend("auto")).name

Why it matters

A worker introduces itself to the hub with its backend name before it starts
taking jobs. Doing that through select_backend("auto") means the machine must
have mlx or torch installed — so a worker started with an explicit
backend_factory raised RuntimeError: No training backend available and died
during registration.

That is the one case backend_factory exists to serve. The parameter was only
half-wired: _run_job already honours the override where the training actually
happens (worker.py:294); registration didn't.

Fallout this fixes

Two tests were failing on main:

- test_full_worker_loop_with_stub_backend
- test_worker_failure_is_reported_not_swallowed

Both hand in a stub backend to exercise the real _Link and the real websocket
without a training stack, and never got past registration. They each also burned
a 15s thread-join timeout waiting on a worker thread that had already died.

- CPU suite: 85 passed / 2 failed → 87 passed
- CPU suite runtime: ~25s → ~5s
- pytest no longer silently requires mlx or torch, which is what CLAUDE.md
already claims of it

Testing

pytest (excluding tests/gpu/, which needs CUDA) — 87 passed.
No new tests: the two that exist covered this exactly, and were failing for it.