Fix OD benchmarks and disallow calculating DetailedPRCurves when using AnnotationType.RASTER

[ntlind]

Improvements

• Fix a bug in the benchmarking script where we weren't actually telling valor_core to calculate DetailedPrecisionRecallCurve
• Throw a NotImplementedError when a user tries to calculate DetailedPrecisionRecallCurve using AnnotationType.RASTER.
  • Currently, valor_core will try to produce (datum_uid, mask) pairs, which easily leads to OOM issues since we're creating sets of tuples, each containing relatively large masks
  • These curves were never intended to be used with rasters (just bounding boxes and polygons), so I'm recommending we just throw a NotImplementedError for now. when we re-prioritize DetailedPRCurves in the future, we can discuss what the right output is for this metric.
• Disable compute_detailed for the Box and Polygon OD benchmarks since a) they make the tests take a while to run (~80 seconds for just one run of 5000 Box datums; 213 seconds for one Polygon run) and b) profiling and improving these PR curves is low priority right now. We can try adding them back after Justin merges his PRs.
• Update the valor_core OD benchmarks to use the ValorDetectionManager class
• Refactor the raster IOU function to make it easier to read