Fix NameError in extract_entities_long; collapse duplicate surfaces when spans are stripped

[blockems]

Two independent fixes to the long-document path, found while serving GLiNER2 behind an HTTP API. Happy to split into separate PRs if you prefer.

1. extract_entities_long raises NameError on every call

extract_entities_long forwards overlap_policy=overlap_policy to extract_long, but its own signature never declares the parameter:

model = GLiNER2.from_pretrained("fastino/gliner2-base-v1")
model.extract_entities_long("some long document ...", ["person"])
# NameError: name 'overlap_policy' is not defined

extract_long and batch_extract_long both declare overlap_policy: Optional[str] = None, so this looks like one wrapper missed when the parameter was threaded through. batch_extract_entities_long is unaffected because it does not forward the name — which is why the batch path works and the single-text path does not.

Fix adds the missing declaration, matching extract_long's default.

2. Duplicate surfaces once spans are stripped

_dedupe_items deduplicates spans by position, which correctly keeps the same surface text at different offsets. When include_spans and include_confidence are both false, _strip_span_metadata then reduces those spans to bare strings — and the distinct positions become identical strings that no caller can tell apart.

On a long document this is very visible. A report mentioning "PostgreSQL" 300 times returns it 300 times:

{"entities": {"database": ["PostgreSQL", "PostgreSQL", ...x300]}}

At that fidelity the repeats carry no information — no offset to distinguish them, no confidence to rank them — so a caller can only discard them.

The fix collapses them while preserving first-seen document order, scoped to lists that reduce entirely to strings. Anything still carrying confidence, offsets or attribute-group payloads is untouched, so include_spans=True behaviour is unchanged.

Verified against fastino/gliner2-base-v1 on a document repeating one sentence 120 times:

mode	before	after
formatted strings	database: ~200 entries	database: 3 unique, location: 1 unique
include_spans=True	49 entries	49 entries, all distinct positions preserved

I read the note in the long-document tutorial about keeping distinct mentions at different document positions — that behaviour is preserved wherever positions are actually visible in the output. If you would rather have this behind an opt-in flag than as a default, say the word and I will rework it.

Testing

tests/inference and tests/processing pass — 71 tests.