Add served model alias support

[krystophny]

Summary

• add a dedicated --served-model-name server flag
• expose a stable alias via the existing served_model_name load path
• cover the new CLI option with parser tests

Why this is deployable on its own

• it only changes how the server advertises and validates model names
• it does not require the Responses branch to remain useful
• it provides stable short aliases like qwen for local clients and benchmarks

Testing

• PYTHONPATH=/Users/ert/code/vllm-mlx /Users/ert/code/.venv/bin/python -m pytest tests/test_server.py -q
• python3 -m compileall vllm_mlx

[krystophny]

This PR is now superseded by upstream waybarrios/vllm-mlx#125, which merged --served-model-name support into main on March 21, 2026:
waybarrios#125

Our fork main is already synced to upstream main, so keeping this PR open no longer improves the stack. I am closing it to keep the fork PR set clean and non-duplicative.