Integrate tile2net task

[annehaley]

This PR adds an orthoimagery (3.6 GB raster) dataset to the Boston floods use case and adds a new simulation type which runs tile2net on a selected imagery dataset. The "Segment Curbs" task downloads tiles of raster data in a format expected by tile2net and then runs the tile2net generate and inference functions. The output files generated by the inference step are then interpreted as new Datasets to add to the current project.

The newly-created vector datasets are sent through our existing vector conversion process and are made available to view on the map. The first Dataset is the "polygons" generated by the inference, which represent the segmentation of curbs, crosswalks, and other walkable paths in the imagery. The second Dataset is the "network" generated by the inference, which is a processed version of the "polygons" set, consisting of semi-connected line segments. An avenue for future work would be to interpret the "network" set into our own Network representation. However, due to the nature of walkable urban systems, the resulting Network would be very disconnected and not viable for meaningful GCC analysis.

Results

• Clifton Park set got poor results, took 2 mins for a 144 MB file (not trained for this area, not very walkable)
  

• Boston set got good results, took 2 hours 10 mins for 3.6 GB file (trained for this area, semi-walkable)
  

• Capitol Hill set got fair results, took 8 mins for a 1 GB file (trained for this area, highly walkable park)
  

[jjnesbitt]

Just left one more comment about the composite_orthoimagery script. Otherwise this LGTM