CoMERA: Computing- and Memory-Efficient Training via Rank-Adaptive Tensor Optimization