PERF: fix performance regression in memory_usage(deep=True) for object dtype

[neilkg]

• closes PERF: Performance regression with memory_usage(deep=True) on object columns #33012
• tests added / passed
• passes black pandas
• passes git diff upstream/master -u -- "*.py" | flake8 --diff
• whatsnew entry

The pull request is to update lib.memory_usage_of_objects from taking self.arrays to self._values. An ASV included to benchmark with and without object-dtype columns.

Before:

After:

[WillAyd]

@jbrockmendel thoughts?

[jbrockmendel]

LGTM. @jorisvandenbossche has stronger opinions than i do on when to use .array

[jorisvandenbossche]

@jbrockmendel This is not about an opinion on when to use .array or not, this is about a cython routine needing a numpy array and not an EA.
(so the alternative to _values would be extract_array(... , extract_numpy=True), but since we already know we are object dtype here, it should be fine to use _values)

[jorisvandenbossche]

Thanks for the PR @neilkg !
Looks good to me

[ShaharNaveh]

@neilkg Can you please merge master?
This should solve the failing test.

---

https://dev.pandas.io/docs/development/contributing.html#updating-your-pull-request

[neilkg]

@neilkg Can you please merge master?
This should solve the failing test.

@MomIsBestFriend done

[mroeschke]

Thanks @neilkg!

[jreback]

this was tagged for 1.04? hmm

not sure if we are doing that but no harm
i guess (though the release note will need to be moved)