Optimizations for UniversalKriging

[rth]

Following explanations in the feature request #2, this PR reintroduces code from v1.0.0 (3bf98a1) for UniversalKriging where it is computed with a loop.

In the version 3bf98a1, in every loop an A matrix is defined together with a second member b and the system is A x = b then solved with np.linalg.solve.

This PR, just factorizes the matrix definition outside of the python loop, and computes it's inverse. The x vector can be obtained by A^(-1) b, which is much faster once we have inverted the matrix. This result in (a) reducing the required RAM, compared to the all vectorized version (b) speed up the code by a factor of ~15 (for nx=100, ny=200, n=100).

Everything is defined in the cexecute function, alongside the original execute function. A typical use would be,

 z, ss = UK.cexecute('grid', gridx, gridy, backend='python')

There are some additional tests in tests.py to check that both versions give the same output (although this might need to be tested more).

Edit:

Added a vectorized pure python version (without the loop), that is ~4-6x faster, but requires more RAM (e.g. 6 GB of RAM for nx=1000, ny=1000, npt=100). It can be accessed through,

 z, ss = UK.cexecute('grid', gridx, gridy, backend='vectorized')