PERF: restore performance for unsorted CategoricalDtype comparison

[qwhelan]

Fixes a performance regression introduced in #26403 very shortly before 0.25.0rc0 was cut, which can be seen here

When comparing CategoricalDtypes with ordered=False (the default), a hash is currently done that is relatively slow even for a small number of categories. If we check if the categories happen to be the same and in the same order, we see a significant speedup in the equal case. The .equals() function pre-exits, so overhead in the non-equal case should be minimal.

asv results:

       before           after         ratio
     [a4c19e7a]       [6810ff2f]
     <unsorted_cats~1>       <unsorted_cats>
-       179±0.4μs       25.6±0.4μs     0.14  indexing.CategoricalIndexIndexing.time_getitem_slice('monotonic_decr')
-         179±2μs         25.2±1μs     0.14  indexing.CategoricalIndexIndexing.time_getitem_slice('monotonic_incr')
-       180±0.4μs       25.1±0.3μs     0.14  indexing.CategoricalIndexIndexing.time_getitem_slice('non_monotonic')

• closes #xxxx
• tests added / passed
• passes black pandas
• passes git diff upstream/master -u -- "*.py" | flake8 --diff
• whatsnew entry

[jreback]

thanks @qwhelan