correct quantile to handle unsorted quantiles

Currently dask.dataframe.core.quantile(df, q) can silently give incorrect results when the list of quantiles, q, is not sorted. For instance quantile(dask.array.arange(100), [0.75, 0.50, 0.25]) gives incorrect results. This patch uses numpy's mergesort to ensure that the quantiles are sorted. Note that with the patch behavior still differs from that in pandas.DataFrame.quantile() where quantiles are calculated correctly while preserving order. While this patch does not duplicate the behavior of pandas because it does not preserve the order of the quantiles, it does at least avoids the silent errors.

dask/dataframe/core.py

@@ -3967,7 +3967,9 @@ def quantile(df, q):
 
     # pandas uses quantile in [0, 1]
     # numpy / everyone else uses [0, 100]
+    # current implementation needs qs to be sorted, sort in-place to make sure
     qs = np.asarray(q) * 100
+    qs.sort(kind='mergesort')
     token = tokenize(df, qs)
 
     if len(qs) == 0: