Collect stats for read/augmentation processors

We need a way to debug performance for invididual processing steps. Frequently training runs are underutilizing GPUs and knowing where the bottlenecks are is critical to speed this up.