Don't use lazy thunks for arguments that are always used

[jespercockx]

Currently compiling code with the --lazy flag incurs a quite heavy performance overhead (e.g. the slowdown is over 10x on Triples.agda). If we could analyze which arguments are always used by a given function, then we could avoid creating delay thunks for those arguments and instead evaluate them strictly. I started implementing this analysis in 2834874, but it is not yet clear to me how to do this analysis properly for mutually recursive functions.