Improve kernel config for fused brute force knn search

[tfeher]

Using the example from #277, one can observe that the performance of the fused distance+knn kernels are not ideal when the number of queries is low. The task is:

• tune the configuration of the fused distance+knn kernels to maximize performance
• re-evaluate the logic to choose between fused and non fused kernels

[tfeher]

Cross referencing FAISS issue https://github.com/facebookresearch/faiss/issues/3621