Add keep_variables keyword to open_dataset()

[claytharrison]

This is based on #895 by @tsupinie, following the suggestions on that PR made by @shoyer, but that was so old I decided it was best to do a completely new PR. XArray has changed quite a bit since then, so I had to make changes in several more files than the original PR did in order to get tests to pass.

I'm very new to working on XArray, and my approach here was to basically just mirror how the package handles drop_variables wherever that appears, with the necessary logical tweaks where appropriate. As such I've possibly touched something that shouldn't be (in particular I'm looking at xarray/backends/zarr.py and xarray/backends/common.py as places where I'm not quite sure if I've done things properly).

• Closes Opposite of drop_variables option in open_dataset() #1754
• Tests added
• User visible changes (including notable bug fixes) are documented in whats-new.rst
• New functions/methods are listed in api.rst

[max-sixty]

Thanks for the PR @claytharrison !

I don't know this code that well; it looks good from my perspective. Do others?

I'm guessing this is most useful if we have hundreds or thousands of variables and only want to load the metadata of a few? Since IME loading the metadata is v fast, and then we can drop the variables in xarray before loading their data?

[shoyer]

I'm not sure we need this new API, or even that drop_variables would make sense if we were considering it today.

Why not open the dataset with decode_cf=False, drop any variables you don't need, and then decode with an explicit call to xarray.decode_cf?