Drop stale funders when acknowledgement overlap cleanup removes their annotation

[Copilot]

Problem

The funding-acknowledgement parser injects inline <rs> entity annotations (funder, grantNumber, …) into acknowledgement text that already contains inline <ref> markup (URLs, citation callouts) from earlier processing. The injector only places an <rs> when its span sits inside a single plain text node, so any <rs> whose span overlaps a <ref> is silently dropped. The structured funding list (<funder> / <org type="funding">) then disagrees with the body: entities are listed with no corresponding inline mention.

Root cause

This is an overlapping-markup problem. XML can represent two elements over the same tokens only when one nests inside the other; the injector doesn't nest — it skips any <rs> that touches a pre-existing inline element. Before this PR, the entity was nevertheless kept in the structured output, leaving the funding list and the inline annotations inconsistent.

Amplifier (fixed separately, upstream): many of these <ref>s were phantom — a pdfalto bug (annotation coordinates referenced to the MediaBox instead of the CropBox) shifted URL link boxes ~one line down onto unrelated text on crop-inset PDFs. Fixed in pdfalto (getMediaBox() → getCropBox()) with its own regression test. But even with a correct pdfalto, a real funder/grant span can legitimately coincide with a real <ref>, so GROBID must keep its output consistent in that case. kermitt2/pdfalto#239

Fix

Keep the structured entity list consistent with the inline annotations by construction, using a direct link between each annotation and the entity it was extracted from:

• AnnotatedXMLElement now carries a reference to its source entity (Funding, Person or Affiliation), recorded when the annotations are built from the labeled clusters.
• The XML-injection methods (updateParagraphNodeWithAnnotations, updateSentencesNodesWithAnnotations) report which annotations were actually injected.
• After injection, pruneEntitiesWithoutInjectedAnnotation removes the entities whose annotations were all dropped. An entity is kept if it has no recorded annotation, or if at least one of its annotations survived.
• Correspondence is resolved by object identity (IdentityHashMap-backed sets), not by (type, text) string matching — two entities with the same textual value (e.g. two fundings citing the same funder name) remain distinct. This replaces the fragile string-matching reconciliation from an earlier iteration of this PR, and covers persons and affiliations as well as fundings.

Trade-off: an entity whose only mention overlaps a legitimate <ref> is dropped rather than annotated. Teaching the injector to nest <rs> and <ref> (both directions are valid TEI) would preserve such entities; that is left as a follow-up.

Tests

• Existing FundingAcknowledgementParser integration tests retained.
• New integration tests using a stub label() implementation:
  • a funder span overlapping a pre-existing <ref> → the annotation cannot be injected and the funding is dropped from the structured output;
  • a funder span in plain text → the annotation is injected and the funding is retained.

Not addressed here (separate follow-ups)

• Nesting <rs>/<ref> in the injector so overlapping mentions are annotated instead of dropped.
• The "unknown funder" symptom (grant/program with no funder → bucketed under Funder.EMPTY in TEIFormatter) — a funding-model/TEIFormatter issue, not markup.