@lorentzenchr would you mind reviewing this PR? If this works, I can use it as a template for the other estimators in sklearn.neighbors. I just multiplied the distance weights by sample_weights to get an overall weight for each point.

One issue I had was with the function _check_sample_weight_equivalence. From the docstring, the function "check(s) that setting sample_weight to zero / integer is equivalent, to removing / repeating corresponding samples." For most estimators this check should pass, but I don't think that it should for the KNN estimator. Duplicating points doesn't just double the weight for a data point it changes how many times that point is calculated as a neighbor. Thus, it shouldn't be the same as doubling the weight of the point.

One way I could see of reconciling this is by removing any duplicate data points before running the predictions and increasing the sample_weight proportionally. Then _check_sample_weight_equivalence would pass. I'm not sure if that is changing the underlying KNN estimator though. It seems like more of a trade off.

I'm also unsure how sample_weight=0 should be handled. If all the samples are 0, an error is thrown as expected. However, there can still be issues when a couple of sample weights are 0. For example, if all k nearest neighbors have a weight of 0, should an error be thrown? Should nan be returned? I think maybe the best thing to do would be to remove these points before getting a point's neighbors.