Unveiling the Backbone-Optimizer Coupling Bias in Visual Representation Learning