Document that using float32 for bounding boxes does not have loss of precision

[kylebarron]

trafficstars

At least, as long as float64 input is provided, so that we can downcast manually

(We should also add tests for this)