This allows building against the same version of GLM across all targets
and not deal with API changes of newer GLM versions which introduce
regressions. Long term we want to migrate away from GLM and stick to a
minimal self-written subset.