COMPAT: .unique() should be type preserving

[jreback]

xref https://github.com/pandas-dev/pandas/pull/15439/files#r101811209

In [1]: Series([1, 2, 2], dtype='int8').unique()
Out[1]: array([1, 2])

In [3]: Series([1, 2, 2], dtype='int32').unique()
Out[3]: array([1, 2])

In [4]: Series([1, 2, 2], dtype='int64').unique()
Out[4]: array([1, 2])

In [5]: Series([1, 2, 2], dtype='float32').unique()
Out[5]: array([ 1.,  2.])

these should all be the original dtype.

We only create hashtables (and use them in unique) for int64, float64 and uint64, should do this for other dtypes.
https://github.com/pandas-dev/pandas/blob/master/pandas/src/hashtable_class_helper.pxi.in#L222

I thought there was an issue about this, but guess not.

[mroeschke]

Looks to be the case on master. Could use a test.

In [82]: In [1]: Series([1, 2, 2], dtype='int8').unique()
    ...:
Out[82]: array([1, 2], dtype=int8)

In [83]: In [3]: Series([1, 2, 2], dtype='int32').unique()
    ...:
Out[83]: array([1, 2], dtype=int32)

In [84]: In [4]: Series([1, 2, 2], dtype='int64').unique()
    ...:
Out[84]: array([1, 2])

In [85]: In [5]: Series([1, 2, 2], dtype='float32').unique()
    ...:
Out[85]: array([1., 2.], dtype=float32)

In [86]: pd.__version__
Out[86]: '0.26.0.dev0+684.g953757a3e'