Suppose I have a dask dataframe with lots of columns, all of them of float dtype. _meta_nonempty creates each column separately and then makes a dataframe out of it. This makes some operations surprisingly slow, such as loc. I see some issues with this:

1. _meta_nonempty can be optimized for some dataframes.
2. loc shouldn't require _meta_nonempty in the first place. It only needs self.obj._meta_nonempty.index which doesn't require recreating all these columns.

import pandas as pd
import dask.dataframe as dd

data = {}
for i in range(10000):
   data["col"+str(i)] = [1.0] * 10
df = pd.DataFrame(data)
ddf = dd.from_pandas(df, npartitions=1)

# any loc on ddf takes several seconds because of slow _meta_nonempty construction
%time dloc = ddf.loc[0]   # 2.82s
%time dmeta = ddf._meta_nonempty    # 2.87s

# loc works near-instantly on the original dataframe
%time loc = df.loc[0]   #  1.69ms