adding eps_t parameter to temporal tools for full 3D IoU calculation

[ramanakumars]

Adding an eps_t which defines the temporal width of temporal shape tools, thereby constructing a fully 3D IoU. This metric uses the displayTime attribute which is a normalized time value for the video, so eps_t is essentially the fraction of the video that two shapes must overlap to be part of the same cluster.

Also updating the solver from scipy.optimize.shgo to scipy.optimize.direct. One main issue was that keeping locally_biased to True was causing poor convergence (particularly on the rectangle tool), but locally_biased=False is definitely much slower.

[hound]

Some files could not be reviewed due to errors:

./panoptes_aggregation/reducers/shape_metric_IoU.py:12: [F401] 'sys' imported...

./panoptes_aggregation/reducers/shape_metric_IoU.py:12: [F401] 'sys' imported but unused

./panoptes_aggregation/reducers/shape_reducer_dbscan.py:16: [F401] 'sys' impo...

./panoptes_aggregation/reducers/shape_reducer_dbscan.py:16: [F401] 'sys' imported but unused

[CKrawczyk]

These changes all look good. The basic idea is to expand each 2D shape into a volume with a width of eps_t in the time direction and calculate the intersection over the union of the volumes. This will preserve the scale independence of this metric and lead to better results.

I have no issue with the optimizer being changed, as long as it is stable I don't care too much about the speed.