BUG: DataFrame(ndarray, dtype=..) does unsafe conversion

[jorisvandenbossche]

From #26848. When you pass a ndarray to the DataFrame constructor and specify a dtype, this does a "plain" numpy astype, which can have some unwanted side-effects (that we avoid in other parts of pandas) such as np.nan -> integer conversion and out of bounds timestamps:

In [26]: pd.DataFrame(np.array([[1, np.nan], [2, 3]]), dtype='int64')
Out[26]: 
   0                    1
0  1 -9223372036854775808
1  2                    3

In [27]: pd.DataFrame(np.array([['2300-01-01']], dtype='datetime64[D]'), dtype='datetime64[ns]')
Out[27]: 
                              0
0 1715-06-13 00:25:26.290448384

Both cases are guarded in DataFrame.astype:

In [29]: pd.DataFrame(np.array([[1, np.nan], [2, 3]])).astype(dtype='int64')
...
~/scipy/pandas/pandas/core/dtypes/cast.py in astype_nansafe(arr, dtype, copy, skipna)
    678 
    679         if not np.isfinite(arr).all():
--> 680             raise ValueError('Cannot convert non-finite values (NA or inf) to '
    681                              'integer')
    682 

ValueError: Cannot convert non-finite values (NA or inf) to integer

In [30]: pd.DataFrame(np.array([['2300-01-01']], dtype='datetime64[D]')).astype(dtype='datetime64[ns]') 
...
OutOfBoundsDatetime: Out of bounds nanosecond timestamp: 2300-01-01 00:00:00

I suppose we want to do such a safe astype in DataFrame constructor itself as well?