Efficient Parameter Optimisation for Quantum Kernel Alignment: A Sub-sampling Approach in Variational Training