fix(go): don't emit embeds for interface type-set constraints

[Synvoya]

Problem

The Go extractor treats every interface type_elem as an embedded interface.
But a Go type-set constraint (generics) is not embedding — a union
A | B or an approximation ~T is a set of type terms, and Go only ever
embeds a lone interface type. Because the branch walked union terms the same
way it walks an embedded interface, each non-predeclared term produced a
spurious embeds heritage edge.

Reproduction on v8 HEAD

package p
type MyInt int
type MyFloat float64
type Number interface {
    MyInt | MyFloat
}

extract_go emits:

Number --embeds--> MyInt
Number --embeds--> MyFloat

embeds is a heritage relation (grouped with inherits / implements /
mixes_in in extract.py), so this asserts a composition/method-promotion
relationship that does not exist — Number cannot embed the concrete types
MyInt/MyFloat; they are constraint terms. This pollutes heritage queries
over the graph with false edges.

Fix

Detect a type-set type_elem (it has a | or a negated_type child — both
tokens appear only in constraint position, never in embedding) and reclassify
its terms as references (context="type_constraint") instead of embeds.
This keeps the type link but drops the false heritage claim. The edge count is
unchanged; only the relation/context is corrected.

Genuine interface embedding is untouched — a lone interface term
(Reader, io.Reader, Sortable[int]) has no |/~, so it still emits
embeds.

Tests

Added to tests/test_languages.py:

• test_go_interface_type_union_is_not_embedding — negative control:
  union terms must not be embeds, and are present as
  references/type_constraint. Fails on HEAD (asserts the spurious
  Number->MyInt embed away), passes with the fix.
• test_go_interface_embedding_emits_embeds /
  test_go_struct_embedding_emits_embeds — positive guards that genuine
  embedding on the existing sample.go fixture (ReaderLogger embeds
  Logger/Reader; DataProcessor embeds BaseProcessor) still works.

python3 -m pytest tests/test_languages.py -q → 317 passed, 13 skipped
(the 13 skips are the optional tree-sitter-dm grammar). No behavior change
outside Go interface type-set constraints.

[Synvoya]

Rebased onto current v8 and revalidated this fix. The focused Go regression tests pass locally, the full local Python 3.12 suite passes (3552 passed, 3 skipped), and the fresh GitHub CI matrix is green on Python 3.10 and 3.12.

This is the strongest, most self-contained fix in my open extractor batch, so I’d appreciate prioritizing this one. I’ll hold off on additional Graphify PRs while the existing batch is reviewed, and I’m happy to adjust or close lower-priority items if you prefer a narrower queue.