Handle content-length of 0

[TomAugspurger]

cc @martindurant.

This test is failing right now. Some layer is truncating the response when the Content-Length is 0, so the assert f.read() == data is failing. f.read() is returning b''. Any suggestions on an alternative test?

It is working for the original issue though

In [1]: import dask.dataframe as dd
df
In [2]: df2 = dd.read_csv('https://fred.stlouisfed.org/graph/fredgraph.csv?id=SP500')

In [3]: df2.compute()
Out[3]:
            DATE    SP500
0     2009-10-23  1079.60
1     2009-10-26  1066.95
2     2009-10-27  1063.41
3     2009-10-28  1042.63
4     2009-10-29  1066.11
...          ...      ...
2603  2019-10-16  2989.69
2604  2019-10-17  2997.95
2605  2019-10-18   2986.2
2606  2019-10-21  3006.72
2607  2019-10-22  2995.99

[2608 rows x 2 columns]

Closes dask/dask#5517

[TomAugspurger]

ping @martindurant if you have suggestions for an alternative test.

[martindurant]

I am puzzled that the pandas load works, but the simple read() doesn't. Would be good to figure out why the two take different paths.

[martindurant]

So this fixes the original case, in which HEAD was giving content length 0, but GET was giving right length. If the content length were 0 in the GET too, I think requests itself doesn't fetch the data.