Context-Parametric Inversion: Why Instruction Finetuning May Not Actually Improve Context Reliance