Adaptive control of reaction-diffusion PDEs via neural operator-approximated gain kernels