More robust testing

[jcrist]

Several issues appeared after the merging of #101. It might be good to think about how to improve test coverage for this library. A few notes:

• Looking at the travis logs, sometimes errors appear but the test still passes (perhaps unintentionally): https://travis-ci.org/dask/knit/jobs/291657224.
• There are several possible code-paths that don't currently have tests (e.g. file names prefixed with hdfs://). It might be good to enumerate the possible cases and try to cover all/most of them.
• test_hdfs_home is always skipped as hdfs3 isn't installed on travis. Is this intentional?

Also might be good to document how a novice developer (myself :)) might get a test environment setup locally.

[martindurant]

I think I skipped the latter point when skimming this before - it is a point well made. This could go in the docs, but there's also the README in continuus_integration/. For myself, I use

docker run -d -v /my/path/code:/code mdurant/hadoop
docker exec -it <hash-from-above> bash