Add dropout to train big model w/ all signals

Also 0 out signals that are missing rather than excluding them completely - ideally make model that can take as many inputs as possible (but knows how to deal with 0 values for rotation / Ti / etc. if they're missing)