Return lazily-encoded QueryOutput from the execute functions

[kalabukdima]

Problem

Worker needs to know the last block that has been written. It also needs JSONL output instead of a JSON array. Modifying the interface to expose that information.

What changed

execute_plan / execute_chunk now return Result<Option<QueryOutput>>. QueryOutput exposes num_blocks / first_block / last_block immediately — block selection completes before any encoding — and encodes one block per write_next_block call. A streaming consumer holds at most one encoded block; buffering callers use into_json_lines(). None means the queried range doesn't intersect the chunk. The handle is fully owned: no lifetime parameter, and it stays valid after the chunk reader drops. The Arrow variants return Option<ArrowOutput> with the same block-range metadata.

Budget-trim bug fixed. The phase-1 scan optimization capped the item scan at the budget cutoff but left the block-header scan uncapped, so the range-end boundary block entered the output with unaccounted weight — inflating both the response and last_block. The header scan is now capped to the same cutoff, matching the non-optimized path.

Worker serializes responses as JSON Lines (matching the legacy engine's wire format — it previously sent a JSON array) and reports the engine's post-trim last_block, falling back to the range end only when the output is empty.

Breaking

Callers of execute_plan/execute_chunk/execute_plan_arrow must drop the writer argument and handle Option. No engine API emits a JSON array anymore.

Draft status

I still need to verify whether the ref cloning change doesn't affect the performance.