Sync with pyRegTab 0.5.1: anchor attribute moves with its values

[shigarov]

Ports the jRegTab 0.5.1 fix for ANCH(n) / REC(n) into the extension by
moving the pinned core to pyRegTab v0.5.1.

What changed upstream

The post-extraction transformation used to apply the permutation only to record
values while returning the schema unchanged. For anonymous attributes that was
invisible; for named ones (produced by AVP) the name-value binding broke —
the column carrying the anchor's name received another attribute's values.
ANCH(n) now moves the attribute itself, name together with its values, so only
the schema order changes.

RTL syntax, the grammar and ATP→RTL serialization are unchanged.

What that means here

The match preview prints the recordset schema verbatim from the core
(preview.rs passes rs.schema.attributes straight through), so patterns
combining AVP with ANCH(n)/REC(n) now render the correct header with no
preview code change.

Nothing in the repository exercised that combination, which is why a wrong
header could ship unnoticed — so this adds a preview test built on pyregtab's
own conformance case semantic/anch_named_attrs rather than invented data.

The test discriminates. Against the v0.5.0 core it fails on the schema
assertion (["Lokaler","Dato","Klasse"] vs ["Dato","Lokaler","Klasse"]) while
the values match on both cores — exactly the binding the fix restores.

Also in this PR

• Hover dictionary regenerated with tools/gen_hover_data.py (not hand-edited);
  the diff is exactly the ANCH and REC(n) entries.
• snippets/rtl.json: the "Settings prefix" description said "anchor column",
  which is now plainly wrong — an attribute moves, a column name is not borrowed.
• grammar/UPSTREAM re-pinned to jRegTab v0.5.1. RTL.g4 is byte-identical at
  v0.5.0 and v0.5.1, so the sha256 line is untouched; the commit sha is
  used rather than the annotated tag object, since the sync check fetches
  raw.githubusercontent.com/.../<commit>/....
• Two highlighting fixtures copied byte-for-byte from the new semantic cases.
  They close real gaps: no fixture carried a genuine <ANCH(n)> setting (only
  <FOO>), and none carried a delimited specification {','} at all.
• Release 0.8.9 + plan under plans/, and a new plans/INDEX.md.

Verification

cargo clippy --all-targets -- -D warnings   clean
cargo test                                  26 unit (was 25) + 3 corpus
npm test                                    check:types + grammar + 39 unit + check:sync
npm run test:e2e                            5 passing
npx vsce package                            114.58 KB (CI limit 200 KB)

No baseline was bent: syntaxes/, grammar/RTL.g4 and server/tests/corpus/**
are untouched, all 11 existing snapshots are byte-identical (two new ones added),
and negative_corpus_positions still passes, so conflict_anch_rec.rtl remains
a compile error under 0.5.1.

Known, recorded as backlog (not fixed here)

• The vendored corpus is stale: 151 positive files vs 304 upstream, and
  server/tests/corpus/VERSION says 2026-07-07 while
  test/grammar/fixtures/CORPUS.md quotes it as 2026-08-26.
• hover_data.rs has no freshness gate at all, unlike RTL.g4.

🤖 Generated with Claude Code