Use TopK node for KNN

[wjones127]

Right now to perform KNN, we compute the top k for each batch, concatenate all the results, and get the top k from those batches. If there are a lot of batches, this can lead to OOM error.

KNN is essentially Project(distance) -> TopK(k=k, order_by=distance), so we might just want to use the DataFusion nodes and build upon them.

There is a tracking issue upstream in DataFusion: apache/datafusion#7195
Also there is a drafted PR for an optimized TopK node: apache/datafusion#7250

We could complete that PR and use that to implement an optimize KNN query plan.

[wjones127]

Also handle here:

lance/rust/lance/src/index/vector/ivf.rs

Line 430 in ff793ad

// TODO: Use a heap sort to get the top-k.