Stop superseded term pages from aborting the build

[Robbie1977]

The problem

Two builds this week died after 93 and 38 minutes on the same class of error:

assemble: ".../ontologies/bfo/BFO_0000001_v6.md:16:1":
  failed to extract shortcode: template for shortcode "alert" not found
assemble: ".../releases/datasets/_v6.md:25:1":
  failed to extract shortcode: template for shortcode "alert" not found

Both are superseded term pages - generator version 6, while version is 9 -
that were never cleaned up. They predate the move off Docsy, so they call
Docsy's alert shortcode, and a missing shortcode template is a hard error, not
a warning. One orphan anywhere in ~763k pages aborts the whole run.

Why the orphans exist

process_term() removed only one generation back:

old_filename = term_id + "_v" + str(version - 1) + ".md"

A term missing from a single run's ID list keeps its older file permanently -
_v6 survives the bump to 7, 8 and 9. They accumulate in blog/ontologies/*
and blog/releases/datasets alike, including one literally named _v6.md, from
a term with an empty short_form.

Orphans are not inert. They are walked and parsed on every build, they can
collide on URL with the current version of the same term, and old enough ones
break the build.

Two commits

Reimplement alert and pageinfo with Docsy's signature and rendering, plus
.alert rules in compat.css using the theme's own tokens so they are styled
rather than bare. pageinfo is pre-emptive: it is the other Docsy shortcode
those pages likely use, and finding out the slow way costs another build.

Prune every superseded version, once per group, with a single directory
scan:

with os.scandir(".") as entries:
    for entry in entries:
        match = _VERSION_SUFFIX.search(entry.name)
        if match and int(match.group(1)) < version:
            os.remove(entry.name)

Not a stat per term per version - that would be millions of extra syscalls
across ~763k terms, and this corpus lives on an NFS volume serving a few hundred
metadata operations per second. Net effect is one fewer stat per term than
before, since the old per-term cleanup is removed.

Verified against the real cases: A_v6.md, B_v1.md and _v6.md removed;
*_v9.md and _index.md untouched.

Note on sequencing

The existing orphans have already been swept by hand. This PR stops them coming
back, and keeps a single missed file from costing a five-hour build.

What this does not fix

The vfb/ directory still holds ~650k files in one directory, which is the
underlying cause of the build's cost. Measured on the NAS: go/ at 14k entries
enumerates at 20,191/s, vfb/ at 650k collapses to 5,076/s and 253 stat/s,
with 128 MiB of page cache available. Sharding by ID prefix is the real fix and
needs its own PR.