Learn from Incomplete Tactile Data: Tactile Representation Learning with Masked Autoencoders