build: Python distribution for 0.2.0, preflight page path, port-note correction

[jiejingzhangamd]

Ahead of tagging 0.2.0. Merge this before the tag — the package metadata fixes have to be in the commit the tag points at.

1. Python package

Nothing in this repo ever built or published one — the only pip invocation in release.yml installs Sphinx, and amd-infera has never appeared on PyPI. The new pypi job builds sdist + wheel, checks the metadata, proves the artifact works, and attaches it to the Release. It needs neither the engine images nor the overlay, so it does not depend on those jobs or gate them.

Which pip packages are involved:

to build	setuptools>=69, setuptools_scm[toml]>=8, wheel — already in build-system.requires, no action
to publish	build + twine — declared nowhere, so the job installs them
the package's own	10 runtime deps plus extras gaie / sglang / vllm / atom / dev

Found by building it before writing the job

fetch-depth: 0 is required. setuptools_scm derives the version from git describe; a shallow clone has no tags, so the wheel comes out 0.0.0 (the fallback_version) instead of the tag. This also means no file needs a version bump — tagging v0.2.0 is what produces 0.2.0.

twine check failed on missing long_description — the PyPI page would have been blank. readme = "README.md" fixes it.

The sdist carried the whole repository. tests (180 files), examples (135), deploy (117), manual (53), bench (45) together outnumbered the source. MANIFEST.in prunes them. The wheel was already clean and is unchanged.

Trial build, with and without this PR

Built twice against a local throwaway v0.2.0 tag (deleted, never pushed):

	version	twine check	sdist
main as-is	0.2.0 ✓	2 warnings	2.3 MB
main + this PR	0.2.0 ✓	PASSED	0.57 MB

Installed into a clean venv both times: version correct, and infera-preflight / infera-kvd-probe / infera-kvd-l3-bench all run.

The tag must be plain PEP 440. A first attempt with `v0.2.0-trybuild` did not degrade — it **killed the build**:

    UserWarning: tag 'v0.2.0-trybuild' version could not be parsed
    ERROR Backend subprocess exited when trying to invoke get_requires_for_build_sdist

So `v0.2.0`, and for a pre-release `v0.2.0rc1`, not `v0.2.0-rc1`.

PyPI upload is opt-in

Skipped unless PYPI_API_TOKEN is set. amd-infera has never been published, so claiming the name is a deliberate decision rather than a side effect of a tag push — and an absent token must not fail the release.

2. Preflight page path

manual/tools/preflight.md was the only Reference entry not under reference/ — my own inconsistency from filing it as "a tool". Moved, toc and inbound links updated. The remaining infera/tools/ strings on that page are source paths, not links.

3. Port-collision note corrected

The GLM-5.2 header recorded "No Address already in use" from the DP-attention run, worded so it reads as "main has this bug and we did not trip it". The run predates #79, which landed the randomised scan start on main (the same fix #59 carried) — the risk is retired, not untriggered. And the observation proved less than the wording implied anyway: that collision needs two engines on one host, which a cross-node PD pair cannot produce. Recording a clean result from a topology that cannot fail the test reads as evidence when it is not.

Audit of the rest, for the record

No stale mixed/ or pd/ path references after the rename; all 7 images referenced from examples/ resolve 200 on Docker Hub; no dead internal links; manual builds with warnings-as-errors. The two infera-overlay:latest strings in deploy/overlay/README.md are a local docker build -t tag, not a registry pull, so the 404 there is expected.