BUG/API: Fix .replace dtype conversion rules

[sinhrks]

Code Sample, a copy-pastable example if possible

1. replace bool with int (#12780)

NG, should be int Series

pd.Series([True]).replace({True: 3})
#0    True
# dtype: bool

2. replace datetime/timedelta with int (#12780)

NG, should be int Series

pd.Series([pd.Timestamp('2011-01-01')]).replace({pd.Timestamp('2011-01-01'): 3})
#0   1970-01-01 00:00:00.000000003
# dtype: datetime64[ns]

pd.Series([pd.Timedelta('1 days')]).replace({pd.Timedelta('1 days'): 3})
#0   00:00:00.000000
# dtype: timedelta64[ns]

3. replace datetime/timedelta with object (#12780)

NG, coerced to internal repr unexpectedly.

rep = {pd.Timestamp('2011-01-01'): 'a', pd.Timestamp('2011-01-02'): 'b'}
pd.Series([pd.Timestamp('2011-01-01'), pd.Timestamp('2011-01-02')]).replace(rep)
#0    1293840000000000000
#1                      b
# dtype: object

4. replace with datetimetz (#12780)

NG, coerced to GMT

pd.Series(['a']).replace({'a': pd.Timestamp('2011-01-01', tz='US/Eastern')})
#0   2011-01-01 05:00:00
# dtype: datetime64[ns]

5. replacement of string (open, see #15743)

In [37]: pd.Series([1,2,3]).replace('1', np.nan)
Out[37]: 
0    NaN
1    2.0
2    3.0
dtype: float64

output of pd.show_versions()

Latest master.

[jreback]

xref (and may close)
#9113

[jreback]

btw, the problem with the datetime with tz is that I force convert=False always (in the block). i don't remember exactly why I did this.

[jreback]

xref #13047

on 32-bit platform some conversions are messed up.

[jorisvandenbossche]

Additional case from #14550:

In [37]: pd.Series([1,2,3]).replace('1', np.nan)
Out[37]: 
0    NaN
1    2.0
2    3.0
dtype: float64

although here it is about coercing of the 'to_replace' value, and not of the fill value.