CLN/PERF: move RangeIndex._cached_data to RangeIndex._cache

[topper-123]

The ._cached_data attribute is not necessary. It was originally added to allow a check to see if the ._data ndarray had been created, but that's also possible to do by a "_data" in _cache check in the new implemention, which IMO would be more idiomatic.

The new implementation has the benefit that the _data will be available to new copies of a RangeIndex, saving the need to create a new ndarray for each new copy of the RangeIndex.

>>> idx = pd.RangeIndex(1_000_000)
>>> idx[[1, 4]]  # this accesses ._data and saves it in cached_data (master) or _cache["_data"](this PR)
>> %timeit idx._shallow_copy()[[1, 4]]
2.55 ms ± 69.3 µs per loop  # master
17.7 µs ± 405 ns per loop  # this PR

xref #26565.

[jreback]

lgtm. @jbrockmendel if you can give a glance

[jreback]

@topper-123 can you rebase

[jbrockmendel]

do we ever do this check outside of tests?

[topper-123]

No, and the checks in the tests are to ensure this array isn't created unless it's needed.

[jbrockmendel]

LGTM pending rebase+green