Determine how to make multiple network outputs flexible

The current export and inference pipeline assumes a single output tensor (detection statistic). Supporting merger time requires routing a second heatmap tensor through the TRT ensemble in export/main.py, receiving and unpacking multiple output arrays in the inference client, and dispatching to the correct postprocessor strategy based on which outputs are present. The goal is to make this configurable (e.g. via export_heatmap: bool) rather than requiring separate code paths.

Start with export/main.py and trace how the export and inference pipeline handles the current detection-statistic tensor. Then inspect the inference client’s output unpacking and the postprocessor dispatch strategy. Done means a configurable export_heatmap option supports routing and processing either one output or both outputs without separate code paths.