PERF: avoid creating numpy array in groupby.first|last

[topper-123]

A unneeded numpy array is created for each group when calling groupby.first and groupby.last on ExtensionArrays. This avoids that.

>>> cat = pd.Categorical(["a"] * 1_000_000 + ["b"] * 1_000_000)
>>> ser = pd.Series(cat)
>>> %timeit ser.groupby(cat).first()
210 ms ± 3.03 ms per loop  # master
78.4 ms ± 766 µs per loop  # this PR

The same speedup is archieved for groupby.last. The above is 3x faster than in master because there are two groups == we save creating two arrays. If there were more groups/larger arrays, we'd get even more improvements.

Also adds some type hints to help understand what parameters these funtions accept.

[jreback]

thanks @topper-123 very nice