PERF: removed coercion to int64 in Categorical.from_codes

[nlee737]

• closes Categorical.from_codes shouldn't coerce to int64 #18501
• tests added / passed
• passes git diff upstream/master -u -- "*.py" | flake8 --diff
• whatsnew entry

[WillAyd]

Since this is performance related - is there an ASV for this? If not can you add and post result?

[nlee737]

Different solution needed; removing coercion causes type errors with reduce operations.

[TomAugspurger]

@fivemok could you post more details about the issues? Future contributors trying to fix this would appreciate it.

[nlee737]

@TomAugspurger apologies for the lack of details. The from_codes function can take arrays of non-numeric types which is coerced to an array of np.int64. Removing coercion completely can break NumPy reduction operations like np.max due to array having a flexible type. I think only removing coercion for arrays of numeric dtype will work. I'll create a new PR after running the tests.

@WillAyd should the full ASV be posted or just for pd.Categoricals?

[WillAyd]

@fivemok just for pd.Categoricals and the relevant method(s)

[nlee737]

@WillAyd @TomAugspurger I created a new PR here for this issue.