Read a completed query job's results by byte offset

Reads a completed job's stored result by byte offset, so a large result can be fetched in pieces and a dropped connection resumed. Typical flow: query/run and query/wait with completionOnly: true until the job is COMPLETE, then call this repeatedly, starting at offset 0 and passing each response's stream_stats.next_offset until it is null. Each result is base64 of a raw byte slice of the Arrow IPC stream; decode and append the slices, then parse the joined bytes. The response has the same text/ndjson framing as query/wait. Results stay readable for about an hour after the job completes; after that, and for a job that has not completed, the job line reports an error status.