PERF: optimize Index.delete for dtype=object

[immerrr]

This should close #6933.

The downside of this patch is that we'll lose type inference that might change index type in rare occasions when index items have different dtypes. This is the same issue that occurred in #6440, and it was ruled out as infrequent and worth dropping for extra performance.

Note though, that the benchmark results are not exactly what I've expected:

In [39]: idx = tm.makeStringIndex(100000)

In [40]: timeit idx.delete(-1)
1000 loops, best of 3: 1.46 ms per loop

In [41]: timeit np.delete(idx, -1)
1000 loops, best of 3: 1.05 ms per loop

The result is faster, but not as fast as I'd expect. This is probably because of object refcounting overhead, because MultiIndex fares a lot better with its Categorial-like implementation:

In [42]: midx = pd.MultiIndex.from_product([[0], idx])

# <..snip..>

In [45]: timeit midx.delete(-1)
1000 loops, best of 3: 225 µs per loop

[jreback]

ok...looks fine...merge?

[immerrr]

Yeah, it's ready.

I wasn't sure about combining release notes, but they did seem relevant. If that's ok, let's merge it.

[jreback]

yeh...combing is fine when pretty related