PERF: apply perf enhancements

[jreback]

added vbenches and make most apply operations a fair bit faster

closes #6013

I can't put this one in the vbench as its cython based, but previously was about 100ms (on my machine)
0.12 clocks in at 15ms (the remaining difference in time is due to the sub-indexing, e.g. x['a'] which in theory could also be speeded but a bit tricky and not a lot more to win)

In [6]: %timeit df.apply(lambda x: integrate_f_typed(x['a'], x['b'], x['N']), axis=1)
10 loops, best of 3: 26.2 ms per loop

Here's a related benchmark from #5654 (a version of which is in the vbenches):

In [16]: s = pd.Series(np.arange(4096.))

In [17]: df = pd.DataFrame({i:s for i in range(4096)})

0.12

In [18]: %timeit df.apply((lambda x: np.corrcoef(x, s)[0, 1]), axis=1)
1 loops, best of 3: 1.11 s per loop

with this PR

In [3]: In [18]: %timeit df.apply((lambda x: np.corrcoef(x, s)[0, 1]), axis=1)

1 loops, best of 3: 1.24 s per loop

-------------------------------------------------------------------------------
Test name                                    | head[ms] | base[ms] |  ratio   |
-------------------------------------------------------------------------------
frame_apply_ref_by_name                      |  15.4320 |  32.2250 |   0.4789 | like the cython example above
frame_apply_user_func                        | 166.7013 | 232.7064 |   0.7164 | bench for example right above
frame_apply_axis_1                           | 108.5210 | 109.1857 |   0.9939 |
frame_apply_pass_thru                        |   6.5626 |   6.4860 |   1.0118 |
frame_apply_np_mean                          |   3.9033 |   0.9290 |   4.2018 | this suffers from some construction overhead, not much I can do
-------------------------------------------------------------------------------
Test name                                    | head[ms] | base[ms] |  ratio   |
-------------------------------------------------------------------------------