Replace internally developed CSR GEMM with a call to MKL

[Vika-F]

• internally developed GEMM was removed
• call to MKL's sparse GEMM was added into clusters assignment step in sparse K-means (see assign_clusters kernel)
• incorrect use of communicator was removed
• Sparse method was aligned with dense method. Now both methods give the same results on the same input data passed in sparse and dense layout respectively