[BUG] neighbors::refine can potentially undo pre-filtering

[achirkin]

When fixing #1793 , I realized that the refinement operation can undo the filtering if the number of valid top-k-results is less than k.

So far, we mark invalid results with infinite distance in IVF-Flat and IVF-PQ. But the neighbors::refine can "refine" the distance back to a small value!
It's not a problem with the fused IVF-PQ similarity kernel, because the invalid results do not make it to the outputs; but the non-fused kernel only marks the bad values with inftinite distance and relies on matrix::select_k to filter out invalid results.
matrix::select_k should normally treat infinity as normal values, so the invalid key-value pairs may make it all the way to the refine operation. Moreover, we may get infinities due to low-precision float overflows; refining these is an expected behavior.

[achirkin]

I think, a reasonably ok-ish solution is to add the filtering template to the refine operation, same as it is passed to ANN search kernels.