SgdEpochTask

implements the SGD optimization loop of one epoch in parallel

The UMAP optimization is a stochastic gradient descent on a shared embedding vector, two different edges may hit the same vertex at the same time. This class applies the Hogwild style parallel optimization (the same as the parallel option of the official python UMAP): the gradient of each edge is applied to the shared embedding vector without any lock.

As a consequence the embedding result is no longer bit-level reproducible between each run, but the convergence of the optimization is not affected at all.