groupby producing incorrect results applying functions to int64 columns due to internal cast to float

[jbezaire]

dupe of #3707

groupby often casts int64 data to floats before applying functions to the grouped data, and then casts the result back to an int64. This produces incorrect results.
Here is a simple example:

import pandas as pd
import numpy as np
label = np.array([1, 2, 2, 3],dtype=np.int)
data = np.array([1, 2, 3,4], dtype=np.int64) + 24650000000000000
z = pd.DataFrame({'a':label,'data':data})
f = z.groupby('a').first()

In [40]: print z
   a               data
0  1  24650000000000001
1  2  24650000000000002
2  2  24650000000000003
3  3  24650000000000004

In [41]: print f
                data
a                   
1  24650000000000000
2  24650000000000000
3  24650000000000004

The result is clearly incorrect and should be

                data
a                   
1  24650000000000001
2  24650000000000002
3  24650000000000004

z.data and f.data are both numpy.int64, which masks the fact that z.data was converted to a float with a loss of precision during the groupby operation.
Tests of software using groupby on int64 will only show a problem if the tests include integers large enough to cause this loss of precision. Otherwise the tests will pass and the software will just quietly produce incorrect results in production.

[jreback]

this is a dupe of #3707. Welcome tests and a fix