Cover reused search-query batch rows with tests

[jnadeau207-collab]

Context

The original version of this PR fixed a stale-search-query-embedding bug in
InputBuffer::new_with_candidates by returning a zero vector instead of an
empty one. That bug was real at the base commit this branch was cut from
(7ba7768), where RankingBatchPrep wrote the reused search-query array with
no prior zeroing.

Upstream 85ac72a landed independently and fixed it directly, by adding
sq_slice.fill(0.0) before the per-item write. That fix is strictly more
complete than the one proposed here: it also clears rows the per-item write
never reaches. The util.rs change is therefore redundant on main and has
been dropped.

What is left

sq_slice.fill(0.0) is currently unguarded by any test. Nothing fails if it is
deleted, and the rows it protects are exactly the ones that are easiest to
overlook:

• items whose input_buffer is None, which the if let Some(..) skips
• rows past .take(length_of_input) when a batch follows a larger one
• items carrying an empty or wrong-width query, which the
  src.len() == search_query_embedding_dim guard skips

Change

Extract the search-query fill from RankingBatchPrep::prepare into
util::fill_candidate_search_query_embeddings so it is reachable from a unit
test without a Python interpreter. The move is mechanical and production
behavior is unchanged.

Five tests cover the rows above, including a query-then-queryless reuse of one
array. Each of them fails if sq_slice.fill(0.0) is removed.

InputBuffer gains #[derive(Default)] so a test can build one with only the
two fields the fill path reads.

[jnadeau207-collab]

Contribution by Jesse Nadeau @villain_thropic on X