Optimize multilabel pair matching

[afermg]

Summary

Optimize find_pairs_multilabel by:

• exploding labels into an inverted relation and equijoining positive pair-label rows
• avoiding the previous all-row list_intersect cross product for positive matching
• projecting only required columns in the negative disjointness query
• preserving pandas and DuckDB relation inputs
• quoting/escaping SQL identifiers
• returning deterministic grouped positive associations

Negative disjointness still requires quadratic candidate comparisons; this PR reduces its projection/materialization cost rather than changing that algorithm.

Benchmark

Real JUMP-lite group_low: 1,342 profiles, 709 labels, 247,644 positive pair-label rows, and 834,814 negative pairs. Two-run medians:

Query	origin/main	This PR	Speedup
Positive	16.072 s	0.193 s	83.1×
Negative	21.627 s	6.802 s	3.18×
Combined	37.699 s	6.996 s	5.39×

Peak process RSS dropped from 5.15 GB to 158 MB. Output cardinalities, label keys, counts, and pair associations matched; this PR deliberately makes pair ordering deterministic, so raw pair-array byte digests differ from the prior query-plan-dependent order.

Absolute times were noisy on a heavily contended host, but the SQL-plan and memory improvements were repeatable.

Validation

• 20 focused multilabel tests passed
• coverage includes pandas/DuckDB inputs, sameby/diffby combinations, repeated/null/empty labels, reserved/spaced/quoted identifiers, ordering, dtypes, and empty outputs
• Ruff and git diff --check passed
• independent follow-up review: OK