Inconsistent implementation of transpose()

[t-hishinuma]

Dense.transpose(A) is Dense = A^T COO.transpose(A) is A = COO^T

correct: A.,transpose() // A = A^T A.transpose(B) // A=B^T

[t-hishinuma]

Performance of transpose() in Dense is low due to poor implementation for square matrix. I refactored it so that it can be parallelized.

I have confirmed that the performance improves with a simple benchmark.