My first thought is to model team goals, and then use a multinomial model to ‘divvy up’ the team’s goals to the players. Is there a reason you want to go in the other direction, so to speak, modeling the players and using that to estimate the teams?
One reason the sum of the expected player-level goals is biased upward is that teams take their foot off the gas when they are leading (see my_blog_post). You could test this hypothesis by looking at the average goals scored by (e.g.) Ramos conditional on how many goals Benzema has.