Interaction terms between categorical and numerical features

Description

While it is possible to create feature interactions with the same individual preprocessing in ColumnTransformer via PolynomialFeatures, I find no (convincing) solution for interactions of features with different individual preprocessing, e.g. categorical column with a continuous numerical column.
Such interactions might improve models from sklearn.linear_model.

Code Example

import pandas as pd
from sklearn.compose import ColumnTransformer
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import OneHotEncoder, PolynomialFeatures


df = pd.DataFrame({'a': ['red', 'red', 'blue', 'blue'],
                   'b': ['high', 'low', 'high', 'low'],
                   'x': [1, 1, 1, 2],
                   'y': [2, 3, 4, 2]
                  })

# interactions for features with no individual preprocessing works fine,
# i.e. numerical ones
column_trans = ColumnTransformer(
    [('xy_num',
      PolynomialFeatures(degree=2, interaction_only=True, include_bias=False),
      ['x', 'y'])],
     remainder='drop')
column_trans.fit_transform(df)

# interactions for ohe encoded also works with helper function
cat_cat = make_pipeline(
    OneHotEncoder(),
    PolynomialFeatures(degree=2, interaction_only=True, include_bias=False)
)
column_trans = ColumnTransformer(
    [('ab_cat', cat_cat, ['a', 'b'])],
     remainder='drop')
column_trans.fit_transform(df)

Expected Results

# no clue for interactions between one-hot-encoded 'a' and 'x' 
column_trans = ColumnTransformer(
    [('a_x',
      magic_pipeline(OneHotEncoder(), 'passthrough'),
      ['a', 'x'])],
     remainder='drop')

Versions

sklearn version 0.21

Start by reviewing the reported ColumnTransformer and PolynomialFeatures examples, focusing on how categorical and numerical columns receive different preprocessing. Determine whether the requested interaction support can be specified from the existing examples, and define completion as a documented or implemented way to create interactions between one-hot-encoded categorical features and numerical features.