Improved result matrix stacking for Hamming GPU implementation

[felixpetschko]

@grst Recently I tried the GPU implementation of the Hamming distance metric with 8 million cells of the Omniscope Covid dataset, which is the largest dataset that is currently available. I noticed performance problems with stacking the blocks of the final result matrix that are computed by GPU. Therefore I implemented a numba function to stack the blocks efficiently - stacking only takes 6 seconds now for 8 million cells with 1 CPU.
With the new stacking implementation I was able to run 8 million cells in ~210 seconds with an Nvidia A30 GPU at the cluster with the GPU hamming metric (~2000 seconds with 64 CPUs with the CPU hamming metric).

[grst]

Hi @felixpetschko,

that's great, thank you!

Would you mind updating the changelog?

LGTM otherwise.