PERF: only apply nanops rowwise optimization for narrow arrows

[jorisvandenbossche]

See the timings in #43311 (comment)

(this is a regression compared to previous release mostly when using ArrayManager (and thus doesn't need a whatsnew) and is covered by the FrameOps benchmarks)

[jreback]

can you add some comments about why these things are needed (e.g. for performance), can ref this issue as well.

[jorisvandenbossche]

Added a comment with reference to my timings and update the docstring

[jorisvandenbossche]

Verified that the remaining failure is only an unrelated ResourceWarning

[jreback]

thanks @jorisvandenbossche

[jbrockmendel]

LGTM. merits an asv?

[jorisvandenbossche]

LGTM. merits an asv?

It gave a huge slowdown in the current ASVs when running with ArrayManager, so in that way it's already captured.

We could add one for wide dataframes as well, but you already need an extreme shape (like 10, 1000), so I personally don't find that a very interesting one to add. Misremembered this, so you actually already see a speed-up with something like (1000, 10).

[jbrockmendel]

so in that way it's already captured

id be more comfortable with one that always runs