PERF: improves performance in remove_unused_categories

[behzadnouri]

on master:

In [1]: np.random.seed(2718281)

In [2]: n = 1 << 20

In [3]: ix = tm.makeCategoricalIndex(n, n // 10)

In [4]: %timeit ix.remove_unused_categories()
1 loops, best of 3: 527 ms per loop

on branch:

In [4]: %timeit ix.remove_unused_categories()
1 loops, best of 3: 216 ms per loop

[jreback]

merged via 7031e5f

thanks you sir!