ShishuLM: Achieving Optimal and Efficient Parameterization with Low Attention Transformer Models