fix(dash): refuse a busy or out-of-range --port before printing the URL

[dchaudhari7177]

Closes #714 (and its PostHog twin #727).

What changes

doberman dash used to print Dashboard: http://127.0.0.1:<port>/?token=… before uvicorn tried to bind. A busy port gave a dead link followed by uvicorn's [Errno 10048], and --port 70000 a traceback ending in OverflowError.

1. --port has min=1, max=65535, so an out-of-range port is a Typer usage error (exit 2). That also covers 0 and negatives.
2. A new _dash_port_is_free(port) does a throwaway bind on (_DASH_HOST, port) and closes it. If the bind fails, dash prints error: port N is already in use, try --port <another> to stderr and exits 1. It doesn't print the URL or call uvicorn.run.
   • The check runs before the heartbeat thread starts, not only before the URL. Otherwise a failed start would touch the heartbeat file and, for about two seconds, route AUTH challenges to a dashboard that never came up.
   • Off Windows the probe sets SO_REUSEADDR, as uvicorn does, so a port only lingering in TIME_WAIT is not reported busy. On Windows the option is left off, because there it would let the probe bind a port that is actually in use.
3. docs/CLI.md's per-command exit-code table gets a dash 2 row (port out of range), and the existing dash 1 row now also covers a busy port.

Tests (tests/unit/test_dash_serve.py)

• test_dash_refuses_a_busy_port_before_printing_the_url holds a listening socket on a port and runs dash --port <that port>. It asserts exit 1, the error line, no Dashboard: line, uvicorn.run never called (monkeypatched), and the heartbeat never touched.
• test_dash_out_of_range_port_is_a_usage_error, parametrized over 0, 70000 and -1: exit 2, no URL, no uvicorn.run.
• test_dash_disables_uvicorn_access_log used --port 0, which the new range check refuses. It now picks a free port.

On main these give 4 failed, 19 passed. On this branch the file passes 23/23, and pytest tests/unit -k "cli_help or docs or dash or exit_code" gives 342 passed, 1 skipped. ruff check and format are clean. Run on Windows / Python 3.12.

One thing I noticed but didn't change: the "Collision audit" paragraph in docs/CLI.md says grep -c "typer.Exit(code=" src/doberman/cli/main.py returns 67, but main already returns 88 (66 × code=1, 20 × code=2). This PR adds one more code=1. Happy to refresh that paragraph here or in a separate docs PR.

AI assistance: I used an AI assistant while writing this change and the tests. I reviewed the diff and ran the tests above myself.