fix(train): move the model to the device before the base evaluation

[FerryQ]

What

Unable to finetune on CUDA or MPS without workarounds.finetune now moves the model to the training device right after load_checkpoint, instead of only inside train_model. One line in laya/train.py, plus a regression test.

Why

Since 0.3.29 (#967), finetune scores the base checkpoint on eval_items before training. The model is still on CPU at that point, but calibration_records sends its inputs to the device, so every GPU fine-tune with calibration items (the default) or --eval crashes before training starts:

• CUDA: RuntimeError: Expected all tensors to be on the same device, but got index is on cuda:0, different from other tensors on cpu
• MPS: RuntimeError: Placeholder storage has not been allocated on MPS device!

CPU is unaffected, which is why the test suite did not catch it. notebooks/laya_finetune_typed_decisions_mps.py hits it too, since it calls finetune(device="mps") with calibration items.

I found it while fine-tuning multilingual on Czech text categorization, on a CUDA VM and on an Apple Silicon Mac. Minimal repro: any finetune(data, "multilingual", out, TrainConfig(epochs=1)) with 10 or more rows on CUDA or MPS.

How it was verified

New test EndToEndTests.test_model_is_on_the_device_before_the_base_evaluation. It records the order of model.to and calibration_records calls, so it catches the bug on CPU: it fails on main and passes with the fix.

python tests/test_train.py         # Ran 54 tests, OK
python tests/test_training.py      # 61 passed, 0 failed (same as main)
ruff check laya/ --select=E9,F63,F7,F82,F401,F811 --line-length=120   # All checks passed
python -m compileall -q laya/ tests/

[NandhaKishorM]

Merged.

Moving the model to the device before the base evaluation is a real fix with a narrow window, and that window is new: #967 added the before-and-after comparison in the last release, which is the only thing that evaluates the base checkpoint before training moves the model. So the bug arrived with the feature, and nothing else would have hit it.

The symptom is a device mismatch on the base pass, which means the before column either crashes or never gets measured, and the whole point of that table is the comparison.