questionFeed: exclude N>1 illustration-vote answers; dedupe contested_card_ids across the view

[WilfordGrimley]

Summary

Two related backend fixes in the What's That Card question feed, sharing files.

1. Printing exclusion now sees N>1 illustration-cluster answers. identify_printing has two
answer paths: a single/unclustered candidate posts to /2/submitPrintingTag/ and writes
CardPrintingTag; a shared-illustration cluster of two or more candidates posts to
/2/submitIllustrationVote/ instead, and cast_illustration_vote writes CardPrintingTag only
when the illustration resolves to exactly one live printing. At N>1 - the exact premise of the
cluster UI that fired - nothing lands on the printing channel, only on CardIllustrationVote.
_voter_answered_printing_card_ids read exclusively from CardPrintingTag, so a voter who
answered via the cluster path cast a real, persisted vote the exclusion couldn't see: the card
stayed UNRESOLVED and was re-served immediately. Production evidence: both of the only two
human illustration votes on record were each followed, within seconds, by an is_no_match
printing vote on the same card - a voter escaping the loop, manufacturing false no-match data.

Fix: _voter_answered_printing_card_ids now unions card ids from CardPrintingTag with card ids
from CardIllustrationVote before the existing md5/phash identity-group widening, so an N>1
answer excludes the card from every printing tier exactly like a direct CardPrintingTag vote
does. No new model or migration - CardIllustrationVote already exists and already carries the
right (card, anonymous_id) grain.

Rejected alternatives (both considered, both wrong for this case):

• Reusing CardQuestionAbstention (PR Differentiate 'Not sure' from 'Skip': record a human abstention (issue #712) #731's "Not sure" record): that model represents a
  human non-answer ("I looked and couldn't tell"); an N>1 illustration vote is a real, weighted
  answer. Conflating the two would corrupt the exact distinction Differentiate 'Not sure' from 'Skip': record a human abstention (issue #712) #731 was built to preserve.
• Writing something on the printing channel from the N>1 branch (e.g. a placeholder
  CardPrintingTag): cast_illustration_vote's own docstring explains why this is deliberately
  never done - the endpoint has no way to express "one of these N printings" without falsely
  asserting an unsupported outcome, corrupting printing consensus. Extending the exclusion query
  instead of the write path leaves that constraint untouched.

2. get_contested_card_ids() deduped across the view boundary. get_next_question_feed_item
and get_remaining_estimate each independently called it - measured at 513-627ms per call
against live production data (single-gunicorn-worker deployment, so this is site-wide latency on
every affected request, not one user's). Following the "compute once, thread as an optional
parameter" convention PR #729 established inside this file, both functions now accept an
optional contested_card_ids parameter (None default - every existing direct caller, including
catalog_stats.py, is unaffected). views.get_question_feed resolves it once and passes it to
both. No cross-request cache introduced.

Test plan

• cd MPCAutofill && pytest cardpicker/tests/test_question_feed.py cardpicker/tests/test_illustration_vote.py cardpicker/tests/test_md5_group_pooling.py -
  150 passed (includes two new regression tests under TestIllustrationVoteAnsweredExclusion
  that fail against pre-fix code: one asserts _voter_answered_printing_card_ids includes the
  card after an N>1 illustration vote, the other asserts a voter is not re-served the card via
  get_next_question_feed_item after casting one)
• cd MPCAutofill && pytest cardpicker/tests/ (excluding the pre-existing, unrelated
  TestSharedCacheNotConfigured/TestGetCatalogStatsView failures in test_catalog_stats.py,
  confirmed identical on unmodified origin/master, and the container-only
  test_federation_hash_tool_parity.py) - 3676 passed, 8 skipped, 0 failed
• pre-commit run (ruff, isort, black, mypy, prettier) on all changed files - clean