fix(tests): de-flake serve tests by isolating ports and failing fast

[bhimrazy]

What does this PR do?

Fixes the recurring flake in tests/test_serve.py, most often seen on test_serve_with_generate_strategy[tensor_parallel]:

AssertionError: Server did not respond as expected. Error: HTTPConnectionPool(host='127.0.0.1', port=8000):
Max retries exceeded ... Connection refused

Test-only change; nothing under litgpt/ is touched.

Root cause

All seven serve tests bind the hardcoded port 8000, and teardown never waits for the server to actually exit — kill_process_tree only sends the signals, and the thread wrapping the non-blocking Popen was joined instead of the process. That leads to three compounding problems:

1. Port still held. A server that hasn't finished shutting down makes the next one exit at bind_socket() via uvicorn's silent sys.exit(1), which the test can only observe as a connection refused.
2. Leaked servers. When an assertion fails, the kill_process_tree call below it is never reached, so the server survives for the rest of the session and takes port 8000 with it. tensor_parallel runs last, so it inherits the debris.
3. No diagnostics. Nothing checked whether the server was still alive, so a crash at second one looked exactly like a slow start, and the failure message said nothing about why.

Changes

• Serve each test on its own OS-assigned free port.
• Poll process.poll() while waiting, and fail immediately with the exit code plus a tail of the server output.
• Always tear down via a context manager, and wait for the full process tree to exit before the next test starts.
• Write server output to a file instead of an unread pipe, which can fill up and block a chatty server (NCCL_DEBUG=INFO on CI) before it binds.
• Fold the boilerplate duplicated across all seven tests into two helpers.

Verification

• ruff check and ruff format --check clean; pytest tests/test_serve.py passes locally (GPU-gated tests skip).
• Fail-fast path confirmed by hand: an invalid CLI flag now reports Server exited with code 2 before it was ready. with the server log, instead of a bare connection refused after 30s.

Left as a draft: the 2-GPU tensor_parallel path can't be reproduced locally, so a few CI runs are the real confirmation.