Replace chunks backend with dask for CSV reading.

[stefanseefeld]

This implements the simple backend replacment. It was tested against dask master. While explicitly specifying a blocksize affects performance, even the default is (now) better than performance with the chunks backend.

This change may obsolete a number of other functions that are no longer reachable, so a little more cleanup may be warranted.
I'm also not quite sure how much testing this needs.

[kwmsmith]

To verify we're doing the lean projection, this needs to be something like:

result = pre_compute(s[s.sepal_length > 5.0].species, csv, comfortable_memory=10)
assert set(result.columns) == set(['sepal_length', 'species'])

Specifically, we need to test the pre_compute() stage explicitly.

[stefanseefeld]

Fixed.

[coveralls]

Coverage decreased (-0.02%) to 89.567% when pulling 21d6632 on stefanseefeld:dask into e601b98 on blaze:master.

[coveralls]

Coverage decreased (-0.02%) to 89.567% when pulling 5b5e024 on stefanseefeld:dask into e601b98 on blaze:master.

[kwmsmith]

We can remove these functions now that we no longer need chunks with CSV:

https://github.com/blaze/blaze/blob/master/blaze/compute/csv.py#L76-L97

Looking at the coverage reports, these functions aren't hit anymore during testing.

[coveralls]

Coverage increased (+0.2%) to 89.741% when pulling b2b3575 on stefanseefeld:dask into e601b98 on blaze:master.

[llllllllll]

what version of pandas is required by dask.dataframe?

[stefanseefeld]

pandas>=0.18.0

[llllllllll]

I was going to say that I am using blaze on a project that supported pandas 0.17, but we just had a talk and decided to drop pandas 0.17 support so that is no longer an issue.

[kwmsmith]

@llllllllll great -- any other comments on this PR? Otherwise will merge.

[llllllllll]

We need to add dask[dataframe] to the requirements, other than that it lgtm