What does this PR do ?

Makes InteractiveLauncher act on an explicit --nproc-per-node instead of
letting the local GPU probe override it. The probe ran first, so the flag lost
to it in two ways:

num_devices = determine_local_world_size(nproc_per_node="gpu")
assert num_devices > 0, "Expected num-devices to be > 0"

if nproc_per_node == 1 or num_devices == 1:
    return self._run_recipe_in_process(recipe_target, config)
else:
    effective_nproc = nproc_per_node if nproc_per_node is not None else num_devices

1. --nproc-per-node N was silently downgraded on a one-GPU host.
num_devices == 1 short-circuits to the in-process path whatever the user
asked for, so automodel config.yaml --nproc-per-node 4 started one worker
and logged "Launching job locally on a single device" without saying the
requested value was dropped. Running several ranks on one device is a normal way
to exercise FSDP/DDP or PP code paths on a dev box, and torchrun supports it.

2. --nproc-per-node 1 failed with no visible CUDA device.
determine_local_world_size(nproc_per_node="gpu") raises when
torch.cuda.is_available() is false, and it ran before the nproc_per_node == 1
branch — the one path that needs no device count, since it runs the recipe
in-process:

ValueError: Cuda is not available.

Also fires under CUDA_VISIBLE_DEVICES="". The error comes from inside
torch.distributed.run, so it names neither the flag nor an AutoModel
requirement.

Scope, to be explicit: this does not make CPU training work — the recipe
still requires CUDA further down. It makes the documented flag behave as
documented and lets the failure come from the framework rather than from a
device probe the requested single-process path never needed.

Changelog

• InteractiveLauncher.launch probes determine_local_world_size only when
  nproc_per_node is None; an explicit value is used as-is.
• The single-process branch keys off the resolved worker count rather than
  nproc_per_node == 1 or num_devices == 1, so an explicit N > 1 reaches
  torchrun even when one device is visible.
• Three CPU tests in tests/unit_tests/launcher/test_interactive_launcher.py:
  explicit 4 on a one-device host reaches torchrun with 4 workers; an explicit
  count never calls the probe; and --nproc-per-node 1 still runs in-process
  when the probe would raise ValueError("Cuda is not available.").

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?

--help already documents the flag as "Number of workers per node for
local/interactive jobs" with no mention of clamping to the device count, so the
code now matches it and no doc change was needed.

Compatibility. Without the flag the behaviour is byte-for-byte the same: the
probe still runs, still asserts > 0, and still picks in-process for one device
or torchrun for more. All 21 pre-existing tests in the file pass unmodified,
including test_interactive_launcher_single_device,
test_interactive_launcher_multi_device, and
test_interactive_launcher_multi_device_explicit_nproc.

Verification (CPU, no GPU needed):

• pytest tests/unit_tests/launcher/ → 103 passed.
• Reverting only interactive.py makes all three new tests fail, so they are
  not vacuous.
• pytest tests/unit_tests/launcher/ tests/unit_tests/_cli/ tests/unit_tests/recipes/ → 1291 passed, 12 skipped.
  (tests/unit_tests/recipes/dllm excluded locally: it fails to import
  transformers.models.diffusion_gemma on main too, unrelated to this change.)
• ruff format / ruff check clean.

Additional Information

• Related to --nproc-per-node is overridden by the local GPU count #3774