A Theory of PAC Learnability under Transformation Invariances