Issue pydata#2921 is about mismatching time units between a time variable and
its "bounds" companion.
However, pydata#2965 does more than fixing pydata#2921, it removes all double
attributes from "bounds" variables which has the undesired side effect
that there is currently no way to save them to netcdf with xarray.
Since the mentioned link is a recommendation and not a hard requirement
for CF compliance, these attributes should be left to the caller to
prepare the dataset variables appropriately if required.
Reduces the amount of surprise that attributes are not written to disk
and fixes pydata#8368.

After encoding these and removing the removal of attributes from time
bounds variables, these should be in the attributes. Removes the checks
that they aren't.

Makes the `time_bounds` calendar really different from `time` and checks
that they are correctly encoded to be non-CF compliant.

for more information, see https://pre-commit.ci