support for units with pint

pint's implementation of NEP-18 (see hgrecco/pint#905) is close enough so we can finally start working on the pint support (i.e. make the integration tests pass). This would be the list of tasks to get there:

• integration tests:
  • implement integration tests for DataArray, Dataset and top-level functions (#3238, #3447, #3493)
  • add tests for Variable as discussed in #3493 (#3654)
  • clean up the current tests (#3600)
  • use the standard assert_identical and assert_allclose functions (#3611, #3643, #3654, #3706, #3975)
  • clean up the TestVariable.test_pad tests
• actually get xarray to support units:
  • top-level functions (#3611)
  • Variable (#3706)
    • rolling_window and identical need larger modifications
  • DataArray (#3643)
  • Dataset
  • silence all the UnitStrippedWarnings in the testsuite (#4163)
  • try to get nanprod to work with quantities
  • add support for per variable fill values (#4165)
  • repr with units (#2773)
  • type hierarchy (e.g. for np.maximum(data_array, quantity) vs np.maximum(quantity, data_array)) (#3950)
• update the documentation
  • point to pint-xarray (see #4530)
  • mention the requirement for UnitRegistry(force_ndarray=True) or UnitRegistry(force_ndarray_like=True) (see https://pint-xarray.readthedocs.io/en/stable/creation.html#attaching-units)
  • list the known issues (see https://github.com/pydata/xarray/pull/3643#issue-354872657 and https://github.com/pydata/xarray/pull/3643#issuecomment-602225731) (#4530):
    • pandas (indexing)
    • bottleneck (bfill, ffill)
    • scipy (interp)
    • numbagg (rolling_exp)
    • numpy.lib.stride_tricks.as_strided: rolling
    • numpy.vectorize: interpolate_na
  • update the install instructions (we can use standard conda / pip now) this should be done by pint-xarray

Review the two unchecked tasks in the issue: making nanprod work with quantities and resolving the type hierarchy for mixed quantity operations. Start with the existing integration tests for DataArray, Dataset, Variable, and top-level functions, then run the relevant test suite. Done means both remaining behaviors work without regressions and the listed integration tests pass.