[FEATURE] Surface prefetch component scores (BM25, cosine similarity) as input to reranking pipelines

Is your feature request related to a problem? Please describe. When executing prefetches or fusion queries, Qdrant obfuscates the scores of the individual components in favor of returning a single, final score in the response.

For example, if I want to execute a hybrid search query with RRF, I need to execute a prefetch on BM25 sparse vector and a prefetch on a dense vector with a similarity measure. However, if the final response presents the RRF rank as the "score" even though this yields little information. For e-commerce and live search applications, which often use second stage rerankers like LTR models, users are forced to execute separate API calls for dense and sparse in order to preserve the individual scores.

Describe the solution you'd like A query-time argument that allows users to opt-in to receiving individual prefetch component scores in the response. In the case of nested prefetches, we should nest the scores similarly.

Describe alternatives you've considered Executing concurrent API calls for each prefetch component and manually fusing the results client-side in order to preserve component score information.

Additional context This has been requested by several customers, and myself when working with hybrid search.