FIX get_layer_device_map StopIteration on CPU-only device maps

[Cunninger]

Fixes #3619

get_layer_device_map() assumed every hf_device_map contains at least one device that is not "cpu" or "disk":

main_device = next(d for d in model.hf_device_map.values() if d not in ["cpu", "disk"])

A valid CPU-only map such as {"": "cpu"} (typical for CPU-only prefix-tuning) raises StopIteration before the existing single-root handler can run.

This keeps the GPU/accelerator path unchanged, and falls back to the first non-disk device (default "cpu") when the map is CPU/disk-only.

Tests

pytest tests/test_integrations.py::TestGetLayerDeviceMap -v

6 related tests in that class: 3 passed locally (CPU-only single-root map, per-layer GPU map, CPU offload mixed with GPU).

AI assistance

AI assistance was used to draft this change. I reviewed every line, reproduced the StopIteration from #3619, and ran the tests above. Coordination comment: #3619 (comment)