Parallel calculation using dask

[miniufo]

Although the current implementation is based on xarray and dask. The solution is first loaded completely into memory. During the multi timesteps calculation, xinvert does not accelerate on a multi-core PC. This is a little strange to me and need some times to make it parallelable.

[miniufo]

The current strategy is that, we chunk along the core dimensions, and use xr.apply_ufunc to invert each chunk. In this fashion, we can make use of dask's parallelism but at a minimum cost of re-factoring the current codes.

[miniufo]

I just realize that to release the power of dask through xr.apply_ufunc, the numba-accelerated ufunc should be annotated using nogil=True. This remove the GIL of python and make use of as many cores as possible in the inversion for multiple time or level datasets.