REF: implement Dtype.index_class

[jbrockmendel]

Allows EAs to specify what Index subclass they want to use. Avoids special-casing pandas-internal EAs.

Small perf improvement

In [2]: dti = pd.date_range("2016-01-01", periods=3, tz="UTC")

In [3]: cat = dti.astype("category")

In [4]: %timeit pd.Index(dti)
6.39 µs ± 80 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each) # <- main
5.69 µs ± 368 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)  # <- PR

In [5]: %timeit pd.Index(cat)
6.37 µs ± 112 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)  # <- main
5.8 µs ± 135 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)  # <- PR

[mroeschke]

Could you make these properties like the base class?

[jbrockmendel]

updated to make both cache_readonly; I think the perf improvement in the OP depends on it

[jbrockmendel]

Updated + greenish

[mroeschke]

Thanks @jbrockmendel