fix: stop re-running the annotation finder per annotation when writing metadata

[daniel-ji]

To be merged before #676.

AnnotationImporter.import_metadata() writes one metadata.json per logical
annotation (identifier), whose files list must include every source that shares
that identifier (e.g. a mask + its oriented-points + a mesh). To find those siblings
it re-ran the global annotation finder and filtered by identifier:

anno_files = [
    item
    for item in AnnotationImporter.finder(self.config, **self.parents)
    if item.identifier == self.identifier
]

The finder is global: one call walks all annotation source blocks, globs
each, instantiates it, and calls is_valid() — which for masks loads + numpy-scans
the full MRC file. So one finder pass reads all source files. import_metadata runs
once per annotation, so the whole pass is repeated N annotation times just to keep the 1–2
siblings of each annotation.

Fix (this PR)

Don't call the finder in import_metadata. Accumulate each annotation's metadata
incrementally as its sources are imported, into a class-level map keyed by the
fully-qualified output path:

identifier_metadata_map: dict[str, dict] = {}   # on AnnotationImporter

def import_metadata(self):
    ...
    meta = self.identifier_metadata_map.setdefault(filename, {...})
    meta["files"].extend(self.get_metadata(path))
    meta["object_count"] = max(meta["object_count"], self.get_object_count(dest_prefix))
    self.annotation_metadata.write_metadata(filename, meta)

Each source adds its own files + object_count to the shared doc and rewrites it;
the last sibling's write contains everything. The finder is no longer called from
import_metadata at all.

Notes / scope

• Single file: ingestion_tools/scripts/importers/annotation.py (~40/33 +/-).
• Moved get_object_count + abstract get_metadata up from BaseAnnotationSource
  to AnnotationImporter (import_metadata now lives on the base and calls them).
• Removed the now-dead written_metadata_files guard. IMPORTANT: do not re-add
  written_metadata_files.append(filename) — with incremental writes it would make
  the 2nd+ sibling early-return and silently drop its files.
• import_metadata sets self.local_metadata = meta so the per-instance accessor
  still reflects the written doc (keeps existing tests' assertions valid).
• Behavior preserved for the "points defined after non-points sources" case
  (object_count = max across siblings, files accumulated) — no longer relies on
  the old FileNotFoundError "wait for all siblings" retry.

Validation

• py_compile passes.
• Local-fs end-to-end tests pass: test_ingest_triangular_mesh / _hff
  (exercise import_item + import_metadata + local_metadata assertions).
• No regression vs baseline in local moto run (7 passed / 39 failed identical with
  and without the change; the 39 are bucket does not exist from an incomplete
  local moto harness, not the change — they pass in CI).
• Traced test_import_annotation_metadata_with_multiple_sources by hand: yields
  object_count == 3, 3 files {zarr, ndjson, mrc}, shapes {Point, SegmentationMask}.