PERF: fix performance for series slices (even more)

[immerrr]

This is a continuation of #6341 and is related to #6370. The latter unfortunately concerns public API and may be not easy to fix gracefully.

So, regardless of the decision to be made in other ticket, this patch will optimize series slicing right away and remove some redundant code (no need to override __getslice__ because Series doesn't inherit ndarray anymore).

[immerrr]

Here's the relevant part of vbench output:

Invoked with :
--ncalls: 3
--repeats: 3


-------------------------------------------------------------------------------
Test name                                    | head[ms] | base[ms] |  ratio   |
-------------------------------------------------------------------------------
series_getitem_pos_slice                     |   0.0510 |   4.7594 |   0.0107 |
frame_drop_dup_inplace                       |   2.4463 |   3.0230 |   0.8092 |
frame_reindex_columns                        |   0.3046 |   0.3696 |   0.8241 |
frame_mult_st                                |   3.8100 |   4.3863 |   0.8686 |
frame_mult_no_ne                             |   3.6847 |   4.2224 |   0.8727 |
<snip>

[jreback]

looks gr8! pls rebase on master and can get this in

[immerrr]

Done.

[jreback]

thank you!