Kernel Sufficient Dimension Reduction and Variable Selection for Compositional Data via Amalgamation