feat: job logs are read and followed by line - #48

Job logs are read and followed by line, on top of tracel-ai/tracel-app#1762 and tracel-ai/tracel-client#61.

tracel jobs logs <n> shows the job's whole log the way kubectl logs does. It reads 1,000 lines at a time from the first one, so output starts right away and a pipe sees every line, and --tail <n> shows only the last lines. --start <byte> is gone.

--follow starts with the last 10 lines, or --tail of them, as tail -f does, so following a long-running job doesn't replay its whole log into the terminal. It then shows each new line as soon as Console receives it, over the job's log websocket. Reading the job and the lines after the newest one every 2 seconds stays the source of truth: a line already shown is dropped, a frame that skips a line is caught up at once by reading after the newest, and a websocket that can't open or is lost leaves following to those reads. tracel run --remote --follow attaches while the job is queued and shows every line from the first.

With --json, every line is a log event carrying its number, stream, timestamp and text, as experiments logs --follow does with its entries, and following ends with the end event. docs/cli.md covers all of it.

This replaces #47, which follows the current websocket: one string per line, with nothing to line it up with the stored log. tracel-client is pinned to the #61 commit until 0.24.0 is released, and the three changes ship together.

I tried it against a dev stack running #1762. A finished job with 150,001 lines prints in full in about 2.5 seconds, --tail and --follow --tail start where they should, and following a running job keeps every line in order.