BUG: numpy bug in np.abs on timedelta64

[jreback]

This is a very subtle bug in how np.abs operates on timedelta64[ns]
It converts them to us for some reason on numpy <= 1.6.2. (maybe numpy 1.7 is better)
workaround is to wrap in a Series and astype

tricky to fix this because we really can't detect is except in finalize, but then
have to check all return dtypes from the operations.....

exhibited also in #2948

In [23]: s1 = pd.Series(pd.date_range('20120101',periods=3))

In [24]: s2 = pd.Series(pd.date_range('20120102',periods=3))

In [27]: s1-s2
Out[27]: 
0   -1 days, 00:00:00
1   -1 days, 00:00:00
2   -1 days, 00:00:00
Dtype: timedelta64[ns]

#### notice this is timedelta64[us] and not ns!
In [28]: np.abs(s1-s2)
Out[28]: 
0   1 days, 00:00:00
1   1 days, 00:00:00
2   1 days, 00:00:00
Dtype: timedelta64[us]

# workaround
In [30]: pd.Series(np.abs(s1-s2)).astype('timedelta64[ns]')
Out[30]: 
0   1 days, 00:00:00
1   1 days, 00:00:00
2   1 days, 00:00:00
Dtype: timedelta64[ns]

[jreback]

going to close this....but handling timedelta64[ns] in general is still somewhat of an issue with generic numpy functions....