Unified common dtype discovery

[sstanovnik]

Common dtype discovery (like np.find_common_type) should be unified into an internal function, with proper handling of pandas dtypes. As of #13917, there are two such implementations:

1. pandas/types/common.py:_lcd_dtypes, which works on numpy dtypes only, but differently than np.find_common_type.
2. pandas/types/cast.py:_find_common_type, which just uses np.find_common_type internally.

Exchanging the first one with the second (which is, judging by the code, more proper), breaks some tests.

Proposed tasks:

• Convert usages of _lcd_dtypes into _find_common_types.
• Extend _find_common_types to properly evaluate pandas dtypes.

Occurences of _lcd_dtypes, the one in internals.py is removed by #13917:

[jreback-~/pandas] grin _lcd_dtype pandas
pandas/core/frame.py:
   47 :                                  _lcd_dtypes,
 3705 :                 new_dtype = _lcd_dtypes(this_dtype, other_dtype)
pandas/core/internals.py:
 4438 :     def _lcd_dtype(l):
 4473 :         lcd = _lcd_dtype(counts[IntBlock])
 4489 :         return _lcd_dtype(counts[FloatBlock] + counts[SparseBlock])
pandas/types/common.py:
  389 : def _lcd_dtypes(a_dtype, b_dtype):

[sinhrks]

_lcd_dtypes is used only in combine and looks buggy (it's called from combine_first):

df1 = pd.DataFrame({'a': [pd.Timestamp('2011-01-01'), pd.NaT]})
df2 = pd.DataFrame({'a': [1, 2]})
df1.combine_first(df2)
#                               a
# 0 2011-01-01 00:00:00.000000000
# 1 1970-01-01 00:00:00.000000002

[jreback]

that might be related to #10567