single point of truth for data and model

Goal: make easier to re-configure for different architectures and datatypes.

• Single file for datatypes
• Config which includes the datatype
• Update train and val entrypoints to use new config
• load_dataset update to use data_config
• directly do a loss function in trainer instead of purpose?

---

Changelog

• main.py sketch no longer needed, ros_adetector covers.