feat(queries): add on_conflict option to enqueue for dedupe conflicts

[janbjorge]

Batch enqueues with dedupe_key fail atomically today: one duplicate raises DuplicateJobError and nothing is inserted. This adds on_conflict: Literal["raise", "skip"] = "raise" to enqueue() across Queries, SyncQueries, InMemoryQueries, and the repository port — the default is unchanged, while skip inserts the non-duplicates and reports the rest.

job_ids = await queries.enqueue(
    entrypoints,
    payloads,
    priorities,
    dedupe_key=dedupe_keys,
    on_conflict="skip",
)
# One entry per input, e.g. [JobId(11), None, JobId(12)]

• Shape-preserving return: skip mode returns one entry per input with None at skipped positions, so callers keep the 1:1 mapping and know exactly what was inserted.
• Enforced in SQL: ON CONFLICT ... DO NOTHING against the partial dedupe index keeps concurrent enqueues correct. The insert reads from UNNEST ... WITH ORDINALITY ordered by input position; align_ids_with_dedupe_keys maps the returned (id, dedupe_key) rows back to input positions in O(n). Benchmarked at parity with the previous query shape up to 100k-row batches.
• Exhaustive branching: every branch on on_conflict ends in assert_never, so a future third mode fails mypy at each decision point instead of silently falling into the raise path.
• In-memory parity fix: the in-memory adapter silently double-inserted within-batch duplicate keys where PostgreSQL raises; it now raises, matching the database.
• CLI: pgq queue gains --dedupe-key and --on-conflict (skip exits 0 with a "Skipped" message, raise exits 1).
• Skipped rows produce no job and no queued log entry. Docs updated: reliability (idempotency), quickstart, CLI reference, completion tracking, ports reference.

Closes #678

Testing

• make check passed (804 passed, 11 skipped; ruff, mypy strict, import-linter)
• New Postgres integration tests: single/batch shape preservation, interleaved null keys (guards the id-to-position mapping against real row contents), within-batch duplicates, no log rows for skipped, concurrent-skip race via pool + gather
• New in-memory and CLI tests; the alignment helper has 12 unit tests including a seeded 1000-case reference-model test that mirrors the ON CONFLICT DO NOTHING semantics independently
• Benchmarked old vs new enqueue query at 1–100k batch sizes (parity; skip adds ~10% for the arbiter index probes)

Checklist

• I have read the Contributing Guide
• I have added or updated tests
• I have updated documentation if necessary