Upgrade of GPU refinement

[nickjcroucher]

• Faster graph refinement with CPU + GPU
• Fixes miscounting of triangles with GPU
• Corrects parsing of refinement command line options

[nickjcroucher]

Thanks very much for the comments, very helpful - mostly addressed now, a couple of remaining questions:

• not sure how the implement the changes to constructNetwork you recommend - sounds like an overloaded function, but I'm not sure how to best do that in python without increasing levels of confusion, or whether you meant something else?
• the G_df dataframe is the main way of speeding up refinement - avoid repetitively generating the edge tuples for each parameter combination - worth allowing both approaches to be used?
• the other speedup is the betweenness_sample - looks like graph-tool uses a faster approximation for calculating this than cugraph, and all the performance increase of the latter is lost unless this subsampling is used. But it is tedious to pass the variable through so many layers of functions, I agree - have put in a question above about what you think the clearest way to do this is.
• I have added the extra information to the docs about GPU use, but this was under installation - I wonder if it might be useful to have a separate page on GPU use?

[nickjcroucher]

I think I'm done with this now, if you're happy to merge?

[johnlees]

Yep sounds good, I will just increase the version and make a quick doc change before merging