optimize k-selection implemention of faiss gpu version

[op-hunter]

Is your feature request related to a problem? Please describe.
no problem

Describe the solution you'd like
improve the performance of k-selection algorithm implemention about 3~5%.
replace the implemention that add element into thread queue's first place by appending the elements
into the queue directly. when the thread queue is full, it will trigger bonic sort.

Describe alternatives you've considered
no.

Additional context
Add any other context, code examples, or references to existing implementations about the feature request here.