BUG/API: timedelta64 series should operate directly with np.timedelta64/pd.offsets

[jreback]

Related #4134 and:

http://stackoverflow.com/questions/18159675/python-pandas-change-duplicate-timestamp-to-unique/18162121#18162121

In [81]: df = DataFrame(dict(time = [Timestamp('20130101 9:01'),Timestamp('20130101 9:02')]))

This is buggy

In [82]: df.time + np.timedelta64(1,'ms')
Out[82]: 
0   2013-01-01 09:01:00.000000001
1   2013-01-01 09:02:00.000000001
Name: time, dtype: datetime64[ns]

In [84]: df.time + np.timedelta64(1,'s')
Out[84]: 
0   2013-01-01 09:01:00.000000001
1   2013-01-01 09:02:00.000000001
Name: time, dtype: datetime64[ns]

Doesn't work

In [83]: df.time + pd.offsets.Milli(5)
ValueError: cannot operate on a series with out a rhs of a series/ndarray of type datetime64[ns] or a timedelta

But this is ok

In [86]: df.time.apply(lambda x: x + pd.offsets.Milli(5))
Out[86]: 
0   2013-01-01 09:01:00.005000
1   2013-01-01 09:02:00.005000
Name: time, dtype: datetime64[ns]

This is ok

In [85]: df.time + timedelta(seconds=1)
Out[85]: 
0   2013-01-01 09:01:01
1   2013-01-01 09:02:01
Name: time, dtype: datetime64[ns]