REGR: groupby sum causing overflow for int8

[phofl]

This fixes the regression from #48018

BUT: Handling this on the cython level like this is really ugly... and will blow up our wheel size if we introduce this everywhere, but I could not find a way to restrict the compile combinations.

If there is no better way of handling this on the cython level I would propose casting values to int64/uint64 before calling the group_sum function. This would introduce an unnecessary cast and reduce performance, hence why I was looking for a solution on the cython level.

cc @jorisvandenbossche cc @mroeschke