Docs need to clarify why one would use a `sampler` func vs `X` (`y`, `sample_weight`)

[PeterDSteinberg]

The docs should explain why there are two main usage workflows supported by fit_ensemble, fit_ea and predict_many:

• Should I use a sampler function to create >=1 sample where a sample is a tuple of X, y, and sample_weight. Some ideas for it:
  • Random sampling that is custom
  • Subsetting data spatially or in time
  • Parallelism
  • Converting a 4-D array to 3-D or 2-D
  • Using xarray.open_mfdataset, e.g. with some arguments for pynio's Grib support
• Or just give a tuple of X, y, sample_weight:
  • Only one file is needed to make the full sample that fits in memory

WIP - I'll fill this out more later