Learning Stochastic Shortest Path with Linear Function Approximation