fix(import): classify label entities when restoring a bank (#3236)

[nicoloboschi]

Fixes #3236.

Found while benchmarking #3208/#3214 against a real production export.

The bug

import_bank_async resolved the target bank's config before import_bank
restored the archive's bank row:

resolved_config = await self._config_resolver.resolve_full_config(bank_id, request_context)
return await import_bank(..., config=resolved_config, ...)

The bank cannot exist at that point — import explicitly refuses to write into an
existing bank — so that resolve only ever sees global + tenant config. The bank's
own entity_labels, which the archive carries in banks.json, arrives later. Retain
Phase 1 then classified every label entity in the imported bank as regular.

Consequences:

• entities: exclude exact-match-only label entities from the trigram fuzzy-match index #3208/perf(entity-resolution): exclude label entities from the trigram fuzzy-match index #3214 was silently inactive on imported banks: the partial index
  WHERE entity_kind != 'label' excluded nothing, so every fuzzy probe kept paying
  the recheck-discard cost the index exists to remove.
• Label values were fuzzy-merged during the import — the exact-match-only path
  (perf(entity-resolution): skip fuzzy probing for exact-match-only label entities #3187) keys off the same classification, so distinct label values could collapse
  into one another (Memory Tags added when Entities are not during retain #1558).
• The migration backfill doesn't cover it: it runs once, and a bank imported
  afterwards has nothing to correct it.

The fix

import_bank now takes resolve_config (a callback) instead of a resolved config,
and calls it right after the bank row is restored. This is deliberately not an
optional extra parameter alongside config: the only correct moment to resolve is
inside import_bank, so making it the only way to supply one keeps the next caller
from reintroducing the bug. import_bank_async is the sole caller.

Measured on a real export

862 documents / 38,162 facts, bank has a free-text brief_bio label group. Imported
into an emptied instance (all other banks dropped — the trigram index has no
bank_id, so a leftover bank pollutes any measurement here):

	pre-#3187 baseline	main as shipped	main + this fix
entity resolution during import	471.75 s	382.61 s	26.01 s
share of import spent resolving	45.8%	40.9%	3.1%
import wall-clock	1030.76 s	934.93 s	850.00 s
entities stamped label	n/a (no column)	0	5,719

18.1× less entity-resolution work than the baseline and 14.7× less than main as
shipped — on this bank the defect was hiding nearly the whole benefit of #3214.

(The 500 label rows visible on unfixed main in the issue came from a post-import
ingest burst, not the import. A post-hoc backfill classified 5,355; the fixed import
produces 5,719 because the broken path had fuzzy-merged some distinct label values
into each other.)

Test

test_bank_import_classifies_label_entities: configure entity_labels, export,
delete, import, assert the label entity returns as entity_kind='label' and a
regular one as 'regular'. Verified it fails on main with 'regular' != 'label'.

test_document_transfer.py, test_admin_bank_transfer.py,
test_repair_bank_vector_indexes.py, test_hermes_templates_import.py,
test_entity_resolver.py and test_entity_labels.py pass; lint.sh + ty clean.