Add lowering for detection post-processing models

@rsuderman is already working on this.

Since Mar 18, 2022.

SSD-based object detection models have a post-processing op that is usually not included in the TFLite model, or is added post-training as a TFLite custom op. This part of the pipeline tends to run on the CPU and is becoming a bottleneck, especially in mobile environments where it is becoming common for the model to run on accelerators.

This is a feature request to add support for the ops used in post-processing so that it too can leverage compiler optimizations and acceleration.

I've attached [this TFLite file](https://github.com/google/iree/files/8298526/ssd_mobilenet_v1_coco_nms.tflite.zip) that includes these ops. When running `iree-import-tflite`, I get the error below:

```
error: The following illegal operations still remain:
tfl.non_max_suppression_v4 (count: 90)
tfl.topk_v2 (count: 2)
tfl.sum (count: 1)
tfl.select (count: 1)
```