Python 3 compatibility

[pksohn]

(This is a very similar writeup to the PR in developer. Pay attention to the first bullet under the code changes section). This PR includes several code changes to be compatible with Python 3, and changes the Travis build to work with Python 3.5 and 3.6. This is necessarily in parallel with changes to developer, pandana, osmnet, and urbansim; running a simulation using Python 3 using this new codebase also requires Python 3 compatibility with several of the above. The Travis build currently grabs the py3-compatible branches of the above, so that will have to be updated over time.

Code compatibility changes

• The main unique change to this repository is a strange behavior: in python 3, when reading in the data file (sanfranpublic.h5) using pd.HDFStore() and registering DataFrames as orca tables, the DataFrame index names (the column name of the column that got set to the index, i.e. df.index.name) get read in as `numpy.bytestypes, then cast to the python 3 nativebytestype. In python 2, they get read in as strings. This surely has to do with the fact that python 3 now treats text and binary data separately, but I'm not sure why this affects the Dataframe index names specifically. Anyway, this makes it such that reading the DataFrames from HDF5 as usual breaks the model system when code tries to refer to the index name, likebuilding_id. So I added some handlers indatasources.py` to encode the index names to string. This is a temporary fix that should have some more robust handling eventually.
• Print statements changed to print functions, along with from __future__ import print_function
• Implicit relative imports changed to explicit absolute imports.
• developer.py module renamed to develop.py to avoid namespace conflict related to absolute imports.
• A couple changes to dict: iteritems() method replaced with items() (same behavior), dict.keys().sort() changed to list(dict.keys()).sort().
• Change to file io (relevant in writing yaml files in test modules)
• True division rather than floor division, with from __future__ import division in python 2. I manually checked all of the division operations against test data and they return the same arrays either way.

Travis build changes

• Now builds on 3 versions of Python: 2.7, 3.5, 3.6
• Conditional installation of pandana from pip (2.7) or github on python3-support branch (along with similar installation of osmnet).
• Installation of developer from python3 branch
• Installation of urbansim from pksohn/python3 fork/branch

Again, we'll just have to keep track of where the Travis builds are pulling code from. It should get much simpler as we merge changes into master branches.

[pksohn]

On that long bullet about the index names: it might just make sense to convert the H5 file via a one-time process and leave the code alone.

[pksohn]

Update: I reverted changes from this commit that converts bytes-type index names upon loading. Instead I upload an HDF5 file that has been converted - I think that's a more sustainable fix.