REF/PERF: Move MultiIndex._tuples to MultiIndex._cache

[topper-123]

Currently, the heavy-to-calculate MultiIndex.values attribute is cached in MultiIndex._tuples. It would be more dogmatic to store it in MultiIndex._cache IMO, which is what this PR does.

This has the added benefit of ._values getting copied over to new copies of the MultiIndex, so also gives a performance boost in cases where copying is needed:

>>> n = 100_000;
>>> df = pd.DataFrame({'a': ['a', 'b'] * int(n / 2), 'b': range(n), 'c': range(20, n + 20)})
>>> mi = pd.MultiIndex.from_frame(df)
>>> mi.values # also caches mi._values in mi._cache
array([('a', 0, 20), ('b', 1, 21), ('a', 2, 22), ...,
       ('b', 99997, 100017), ('a', 99998, 100018), ('b', 99999, 100019)],
      dtype=object)
>>> %timeit mi._shallow_copy().values
22.8 ms ± 3.43 ms per loop  # master
34.6 µs ± 997 ns per loop  # this PR

[jreback]

lgtm. can you rebase as merged your other MI PR; ping on green.

[topper-123]

The CI complained that this doc string element is in a different order than in the signature, so I moved it to be after mode like in the signature, which must be correct.

I got no idea why this is failing here and not in other PRs...

[topper-123]

Rebased.

[jreback]

thanks @topper-123