Avoid storing full distance matrix during validation

[AlekseySh]

We don't really need to store full QxG matrix due to its weight.

We can compute matmul + top_k in a batchwise rejime and then clip the second dimension, so the resulted matrix is QxC, where C >> k (where k is from map@k, precision@k, recall@k)

[AlekseySh]

now it's a part of: #522