Allow reusing weights saved in a pyramid during xESFM regridding

[andersy005]

This is my attempt at improving pyramid_regrid's performance and circumventing the ESMF errors when using multithreaded dask workers.

• Creating weights from multiple threads/processes fails (ESMC_GridCreateNoPeriDim)  pangeo-data/xESMF#141

This PR introduces two new functionalities

• a helper function for generating a weights pyramid. with this function, we should be able to serially generate weights without worrying about ESMF parallelism issues.
• pyramid_regrid() accepts a new argument (weights_pyramid). with the pre-generated weights, i imagine that it's possible to turn existing loop within pyramid_regrid() into an embarrassingly parallel computation in the future.

In [1]: import xarray as xr, dask

In [2]: from ndpyramid.regrid import generate_weights_pyramid, pyramid_regrid

In [3]: ds_in = xr.tutorial.open_dataset('rasm', chunks={'time': 12})

• weights generation

In [6]: levels = 4

In [7]: %%time
   ...: with dask.config.set({'scheduler': 'sync'}):
   ...:     weights_pyramid = generate_weights_pyramid(ds_in.isel(time=0), levels)
   ...: 

CPU times: user 3min, sys: 12 ms, total: 3min
Wall time: 3min

• generating pyramids with pre-generated weights

In [8]: %%time
   ...: pyramida = pyramid_regrid(ds_in, levels=levels, weights_pyramid=weights_pyramid)
   ...: 
   ...: 
CPU times: user 32.6 s, sys: 32.1 ms, total: 32.6 s
Wall time: 32.6 s

• generating pyramids without pre-generated weights

In [9]: %%time
   ...: pyramidb = pyramid_regrid(ds_in, levels=levels, weights_pyramid=None)
   ...: 
   ...: 
CPU times: user 2min 59s, sys: 0 ns, total: 2min 59s
Wall time: 2min 59s

Cc @jhamman, @norlandrhagen