PERF: Avoid materializing values in `Categorical.set_categories`

[TomAugspurger]

In Categorical.set_categories, we allocate an array of the values, which may be expensive:

pandas/pandas/core/categorical.py

Line 774 in 83436af

values = cat.__array__()

It should be possible to do this operation by just manipulating the codes.

In [6]: c = pd.Categorical(['a'] * 100000)

In [7]: c.set_categories(['a', 'b'])
Out[7]:
[a, a, a, a, a, ..., a, a, a, a, a]
Length: 100000
Categories (2, object): [a, b]

See 5ab0123 for how this might work, which will probably be squashed, but it's the implementation of Categorical._set_dtype in #16015

I may get to this as a followup to that PR.