Clean up search execution and add optional query CPU pool

[antonchen]

Summary

• clean up the search handlers around a shared prepared-search execution path
• add an optional query CPU worker pool so CUDA ingest can stay separate from query and rerank encoding
• keep the change focused on search execution behavior and model-pool routing

Validation

• cargo check -p next-plaid-api --features "openblas model"
• cargo clippy -p next-plaid-api --all-targets --features "openblas model" -- -D warnings
• cargo test -p next-plaid-api --lib --features model query_on_cpu_routes_logical_lanes_to_distinct_pools -- --nocapture
• cargo test -p next-plaid-api --lib --features model build_search_request_config_rejects_zero_top_k -- --nocapture
• cargo test -p next-plaid-api --lib --features model map_keyword_search_error_marks_malformed_match_as_bad_request -- --nocapture

Implemented entirely by Codex.