BUG: inconsistent upcasting dtypes between diff and shift for int8/int16

[lukemanley]

Pandas version checks

• I have checked that this issue has not already been reported.

• I have confirmed this bug exists on the latest version of pandas.

• I have confirmed this bug exists on the main branch of pandas.

Reproducible Example

import pandas as pd

pd.Series([1, 2, 3], dtype="int8").agg(["diff", "shift"]).dtypes

Issue Description

It appears diff and shift have inconsistent upcasting of types when the input is either int8 or int16:

diff     float32
shift    float64
dtype: object

Its possible there is good reason for this, but I couldn't find it documented or explained anywhere.

Expected Behavior

More consistency in upcasting between diff and shift.

Installed Versions

INSTALLED VERSIONS

commit : a221491
python : 3.8.12.final.0
python-bits : 64
OS : Windows
OS-release : 10
Version : 10.0.17134
machine : AMD64
processor : Intel64 Family 6 Model 158 Stepping 10, GenuineIntel
byteorder : little
LC_ALL : None
LANG : None
LOCALE : English_United States.1252

pandas : 1.5.0.dev0+133.ga221491f99.dirty
numpy : 1.21.5
pytz : 2021.3
dateutil : 2.8.2
pip : 21.3.1
setuptools : 60.5.0
Cython : 0.29.26
pytest : 6.2.5
hypothesis : 6.35.0
sphinx : 4.3.2
blosc : None
feather : None
xlsxwriter : 3.0.2
lxml.etree : 4.7.1
html5lib : 1.1
pymysql : None
psycopg2 : None
jinja2 : 3.0.3
IPython : 8.0.0
pandas_datareader: None
bs4 : 4.10.0
bottleneck : 1.3.2
fastparquet : 0.7.2
fsspec : 2021.11.0
gcsfs : 2021.11.0
matplotlib : 3.5.1
numba : 0.55.0
numexpr : 2.8.0
odfpy : None
openpyxl : 3.0.9
pandas_gbq : None
pyarrow : 6.0.1
pyreadstat : 1.1.4
pyxlsb : None
s3fs : 2021.11.0
scipy : 1.7.3
sqlalchemy : 1.4.29
tables : 3.6.1
tabulate : 0.8.9
xarray : 0.18.2
xlrd : 2.0.1
xlwt : 1.3.0
zstandard : None

[lukemanley]

It looks like the diff upcasting to f32 is done here:

pandas/pandas/core/algorithms.py

Lines 1605 to 1613 in 2db3b0a

	elif is_integer_dtype(dtype):
	# We have to cast in order to be able to hold np.nan
	# int8, int16 are incompatible with float64,
	# see https://github.com/cython/cython/issues/2646
	if arr.dtype.name in ["int8", "int16"]:
	dtype = np.float32
	else:
	dtype = np.float64

and the docstring for that function actually mentions diff and shift:

pandas/pandas/core/algorithms.py

Lines 1543 to 1546 in 2db3b0a

	def diff(arr, n: int, axis: int = 0):
	"""
	difference of n between self,
	analogous to s-s.shift(n)

[jorisvandenbossche]

Sidenote: "diff" and "shift" are not aggregations, and so calling them in agg is a bit strange (IMO we should actually disallow this?)

In [80]: pd.Series([1, 2, 3], dtype="int8")
Out[80]: 
0    1
1    2
2    3
dtype: int8

In [81]: pd.Series([1, 2, 3], dtype="int8").agg(["diff", "shift"])
Out[81]: 
   diff  shift
0   NaN    NaN
1   1.0    1.0
2   1.0    2.0

[rhshadrach]

@jorisvandenbossche

(IMO we should actually disallow this?)

#35725 - trouble is that apply/agg code paths are intertwined in various places. This is one of the main motivations behind #41112.