BUG: Accept column indices for na_values in read_csv

[gfyoung]

>>> from pandas.compat import StringIO
>>> from pandas import read_csv
>>> data = 'a\nfoo\n1'
>>>
>>> read_csv(StringIO(data), na_values={0: 'foo'}, engine='c')
...
TypeError: Expected list, got set
>>> read_csv(StringIO(data), na_values={0: 'foo'}, engine='python')
     a
0  foo  # Should be NaN
1    1

This behaviour is slightly inconsistent with what we do with usecols for example, so it would be nice to be able to process column indices with na_values too.

xref #7119.

[gfyoung]

Reclassifying as a bug because the doc makes it sound like that this behaviour should be possible.