`SimpleImputer.fit` strange behavior with `median` and `most_frequent` strategies

**What happened**:
`SimpleImputer.fit` with `median` and `most_frequent` strategies on frames compute different results comparing to `scikit-learn`.

**What you expected to happen**:
They should have consistent results with `sklearn.impute.SimpleImputer`.

**Minimal Complete Verifiable Example**:

```python
df = pd.DataFrame({"A": [1, 1, np.nan, np.nan, 2, 2]})
# This should return the smallest value
b = dask_ml.impute.SimpleImputer(strategy="most_frequent", fill_value=None)
b.fit(df)
b.statistics_
>>> A 2.0
>>> dtype: float64

c = sklearn.impute.SimpleImputer(strategy="most_frequent", fill_value=None)
c.fit(df)
c.statistics_
>>> array([1.])
```

With `median`:
```python
df = pd.DataFrame({"A": [1, 1, np.nan, np.nan, 2, 2]})
df = dd.from_pandas(df, 2)
b = dask_ml.impute.SimpleImputer(strategy="median", fill_value=None)
b.fit(df)
b.statistics_
>>> A 1.0
>>> dtype: float64

c = sklearn.impute.SimpleImputer(strategy="median", fill_value=None)
c.fit(df)
c.statistics_
>>> array([1.5])
```

**Environment**:

- Dask version: 2021.01.1
- Python version: 3.7.6
- Operating System: MacOS
- Install method (conda, pip, source): pip