KeOps for huge amount of input data points

[javoha]

Is your feature request related to a problem? Please describe. GemPy has difficulties handling huge amount of input data points. The underlying cokriging algorithm and the resulting huge matrices cause memory issues.

Describe the solution you'd like Suggestion by @AndrewAnnex to use the KeOps library instead of the Theano implementation.

Describe alternatives you've considered Working with sparse matrices (#276) or point reduction. Domain subdivision might also offer a solution in some form. On a different level: We might need to think about the universal cokriging just not being the optimal interpolator in cases where dense and area-wide data is available.

Additional context For the integration of pymc4 Theano and Tensorfow are required. Models that do not require a probabilistic framework could be handled differently.