Quick shift fixes

[dalejung]

Adjustments to the quick fix PR.

https://gist.github.com/dalejung/9731798#file-vb_suite-log

-------------------------------------------------------------------------------
Test name                                    | head[ms] | base[ms] |  ratio   |
-------------------------------------------------------------------------------
frame_shift_c_order_axis1                    |  14.4524 |  52.8037 |   0.2737 |
frame_shift_c_order_axis0                    |  16.9516 |  46.5320 |   0.3643 |
frame_shift_axis0                            |  15.2656 |  16.3440 |   0.9340 |
frame_shift_axis1                            |  22.2600 |  22.0567 |   1.0092 |

Note, that the speed up was really in sending a C order array to np.take. You'll notice that master is slow for frame_shift_c_order_ which are tests where the block.values is c_contiguous. Thus current master always transposes.

The original example had it's speed up because it np.rolled df.values which is C ordered (fast) while the .blocks were were (slow).

[jreback]

good catch on this
thanks!