A Large-Deviation Analysis of the Maximum-Likelihood Learning of Markov Tree Structures