Keep model parameters for fuse GGUF export

[abnormal749]

Summary

• Forward donate_model from utils.save() to save_model() instead of always donating the model.
• Add regression coverage for both utils.save(..., donate_model=...) behavior and the mlx_lm.fuse --export-gguf path.

Why

utils.save() accepted a donate_model argument but ignored it and always called save_model(..., donate_model=True). That clears the in-memory model parameters even when callers explicitly need them after saving.

mlx_lm.fuse --export-gguf passes donate_model=False because it saves the fused model first, then reads model.parameters() to export GGUF weights. When the flag is ignored, GGUF export receives empty model tensors.

Related to #353, specifically the path where the generated GGUF is missing expected tensors. This does not claim to cover the separate row-major serialization symptom mentioned there.

Validation

• python -m pre_commit run --files mlx_lm/utils.py tests/test_utils.py tests/test_gguf.py
• python -m unittest -v tests.test_utils.TestUtils.test_save_respects_donate_model tests.test_gguf
• Real-world smoke test with mlx-community/Llama-3.2-1B-Instruct-bf16: trained a one-step LoRA adapter, ran mlx_lm fuse --export-gguf, and confirmed the produced GGUF contains token_embd.weight.