BUG: DataFrame and datetime scalar ops may results in incorrect dtype

[sinhrks]

Code Sample, a copy-pastable example if possible

Timestamp subtraction resuts in Timedelta

pd.Timestamp('2011-01-01') - pd.Timestamp('2010-01-01')
# Timedelta('365 days 00:00:00')

But it doesn't when broadcasted to DataFrame. It's because DataFrame op calls
_try_corce_results which coerces the result to datetime64.

df = pd.DataFrame({'a': [pd.Timestamp('2011-01-01')]})
df - pd.Timestamp('2010-01-01')
#            a
#0 1971-01-01

Expected Output

#          a
#0 365 days

output of pd.show_versions()

current master

[jreback]

dupe of #12437