Fix Chronos-2 fine-tuning to preserve loaded GPU device index

[dario-fumarola]

Summary

• override Chronos2Trainer._move_model_to_device to preserve the model's existing CUDA device when the model is already loaded on a specific GPU and no hf_device_map is set
• this prevents transformers.Trainer from moving a model loaded on e.g. cuda:5 back to cuda:0
• add dedicated unit tests for the CUDA-preservation behavior and guard cases

Why

Issue #457 reports that Chronos-2 fine-tuning fails when the model is loaded on a non-zero GPU index because the default Trainer move logic forces the model onto cuda:0.

Validation

• uv run --python 3.11 python -m pytest test/test_chronos2_trainer.py
• uv run --python 3.11 python -m pytest test/test_chronos2.py -k "pipeline_can_be_finetuned or two_step_finetuning_with_df_input_works"
• uv run --python 3.11 mypy src test

Fixes #457