Fine-Tuning Strategies for Preserving In-Context Learning in Linear Attention