PERF: vectorized DateOffset with months

[chris-b1]

This is a follow-up to #10744. In that, vectorized versions of some offsets were implemented, mostly by changing to periods and back.

The case of shifting by years/months (which is actually most useful to me) required some extra hoops and had poorer performance - this PR implements a special cython routine for that, for about a 10x improvement.

In [3]: s = pd.Series(pd.date_range('1900-1-1', periods=100000))

# Master
In [4]: %timeit s + pd.DateOffset(months=1)
1 loops, best of 3: 140 ms per loop

# PR
In [2]: %timeit s + pd.DateOffset(months=1)
100 loops, best of 3: 14.2 ms per loo

[max-sixty]

This will lose the name of the Index - I think you want to use _shallow_copy:

In [57]: df.index._constructor(df.index)
Out[57]: 
Int64Index([0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18,
            19],
           dtype='int64')

In [58]: df.index._shallow_copy(df.index)
Out[58]: 
Int64Index([0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18,
            19],
           dtype='int64', name='hi')

(there's a broader issue about the constructor keeping the name)

[chris-b1]

It actually doesn't cause a problem in this case because the values are going to be unboxed/boxed here anyways:
https://github.com/pydata/pandas/blob/master/pandas/tseries/index.py#L716

But I'll change it if it's the more idiomatic way to get the constructor?

[jreback]

yes shallow copy is the idiom

[chris-b1]

@jreback - pushed changes for your comments and added MonthEnd and MonthBegin

For your comment about using this routine in the DateOffset are you talking about cythonizing the actual apply method too? On that front, I see there is an offsets.pyx file, but doesn't look it's actually used?

[jreback]

can you add a doc-string to these

[jreback]

• do we have an asv benchmark for this? if not pls add one. otherwise minor change. lmk
• add to the perf section (e.g. just add this issue on in the whatsnew next to Allow DateOffset addition with Series #10744)

[chris-b1]

Made those doc changes. Yeah, there is an asv for DateOffset

    before     after       ratio
  [8ea8968 ] [c1d53f9 ]
-  121.83ms    18.49ms      0.15 timeseries.timeseries_datetimeindex_offset_fast.time_timeseries_datetimeindex_offset_fast
-  121.50ms    21.70ms      0.18  timeseries.timeseries_series_offset_fast.time_timeseries_series_offset_fast

[jreback]

@chris-b1 thanks! these pr's are awesome! keep em coming!