inconsistent types resulting from Timestamp arithmetic

[rosnfeld]

Adding vs subtracting a datetime.timedelta from a Timestamp yields different return types:

In [1]: timestamp = pd.Timestamp(datetime.datetime(2014, 3, 4))

In [2]: timestamp + datetime.timedelta(seconds=1)
Out[2]: Timestamp('2014-03-04 00:00:01', tz=None)

In [3]: timestamp - datetime.timedelta(seconds=1)
Out[3]: datetime.datetime(2014, 3, 3, 23, 59, 59)

It's an easy fix, I have the code change done and I am just polishing up additional tests to document this behavior and ensure nothing else is broken.

In #4547 I commented I was looking at another issue but hadn't created an issue for it, this is that issue. I realized it's probably better to have an issue, for release notes and the like.

[jreback]

yep....1 issue per PR makes things simple, but whatever works