Modify naive algorithm to PAM. Fixes #1

[mailology]

I have modified the naive algorithm to PAM. The logic of swap_naive is changed so that the best and second best distance of each datapoint are cached and used in the computing the cost. When we swap medoid m_k with non-medoid x_i, the cost for datapoint x_j is computed as

1. if x_j is assigned to medoid m_k: cost = min(dist(x_i, x_j), second best distance of j)
2. if x_j is not assigned to medoid m_k: cost = min(dist(x_i, x_j), best distance of j)

Hence, the complexity of PAM is O(kn^2) . The algorithms are tested on the MNIST-1k data as follows:

Original naive algorithm

PAM algorithm

[mailology]

I have fixed the following:

• Need to clarify that the updated algorithm does not use FastPAM1 optimization. FastPAM1 will be implemented in a separate issue. The updated algorithm is PAM.
• Edited some comments to make the code more readable.
• Fixed some coding style to make the logic clearer.
• Run further examples to test the running time and order of k in the complexity as follows. Both PAM and naive give same set of medoids. The order of k for PAM is around 0.67 while that of naive is around 1.7.

PAM algorithm

naive algorithm