This presupposes that classrooms can be thought of as random samples of the child population. We know that they are actually geographically localized samples, so that presupposition is somewhat strained and needs a bit of defending, preferably by empirical means.

Obviously a given class is only a random sample from the population of a given school. So if you want to evaluate a teacher, all you need to do is compare his/her class to other classes from the same geographic region.

Or even better, you can compute a statistical predictor of his student's expected performance, and measure the deviation between predicted and actual performance.

The point is that individual variation in students is irrelevant. It is extremely unlikely that a teacher will get a class full of dumb kids.