Don't drop unreduced variables

[malmans2]

• Closes ds.mean('dim') drops strings dataarrays, even when the 'dim' is not dimension of the string dataarray #5368
• Tests added
• Passes pre-commit run --all-files
• User visible changes (including notable bug fixes) are documented in whats-new.rst

Reduce methods such as mean drop non-numeric variables. However, there's no need to drop variables that are not actually reduced (i.e., when "reduce_dim" not in da.dims).

cc: @rcaneill

[max-sixty]

Great! Thanks @malmans2 !

[max-sixty]

I'll leave this open for a day or so, as this is the sort of area where I'm going to miss some intricacy that someone else is familiar with.

[keewis]

thanks, @malmans2, looks good to me (it took me some time to really figure out how that condition works, though).

[keewis]

actually, wouldn't it make more sense to add unreduced variables without calling reduce? Something like

if not reduce_dims:
    variables[name] = var
elif not numeric_only or np.issubdtype(var.dtype, nb.number) or var.dtype == np.bool_:
    # ...
    variables[name] = var.reduce(...)

[malmans2]

That was my first idea, but a few test failed, so it would be a breaking change. I think it mainly has to do with how attributes are handled by reduce.

[keewis]

well, attrs seems to be one of the issues (but one we could fix!). However, some reduction functions (e.g. std, var) need to run on variables that don't have the reduce dims, so my suggestion is actually wrong. See also the concept of invariant_0d introduced in #5207.

As I'm sure I will have forgotten about this in a few months: could you add a comment explaining this?

[malmans2]

Done!

[keewis]

Great, thanks!

[keewis]

last question: if I'm understanding this correctly, not numeric_only and all following lines will be overridden by not reduce_dims. Should we move that three lines down, below the dtype checks?

[max-sixty]

@malmans2 would you like to add a whatsnew?

[max-sixty]

@malmans2 thank you and please feel free to add a whatsnew in another PR!