fix(fts): preserve exact top-level multimatch ties

[BubbleCal]

What is the bug?

Top-level cross-column MultiMatch pushes the global k into each field's Match child. The leaf path truncates score ties before final row IDs are resolved, so the final MAX(_score) aggregate and (score DESC, row_id ASC) sort can only reorder the wrong survivor set.

On the 10M-row MMLB direct workload, current main mismatched the exhaustive oracle in 121 of 500 cases and produced unstable timed signatures for the same boundary-tie queries.

Linear: https://linear.app/lancedb/issue/OSS-1602

How does this PR fix the problem?

• Plans one bounded, exact single-column CompoundQueryExec per field.
• Keeps field scoring independent and retains the existing bounded outer DisMax / MAX(_score) aggregation.
• Plans each field independently: complete, unchanged indexed fields retain an exact bounded compound scorer.
• Replans only partial-index, overlay-backed, missing-index, or otherwise ineligible fields through the unbounded exact leaf fallback before the final aggregate.
• Preserves exhaustive execution for unbounded queries.
• Supports both explicit MultiMatch and fieldless Match expansion.
• Canonicalizes signed zero boosts to +0.0 before field expansion so MAX aggregation and total-order sorting agree.

The generic cross-column compound scorer was also prototyped for this root shape, but it paid row-address merge costs without a selective probe leaf. The field-local exact design was faster and keeps the final intermediate set bounded by fields * k.

Correctness validation

Environment: 10M-row MMLB dataset, two independently indexed text columns, c4-highmem-16, 8 workers, 64 GiB index cache.

Oracle case	Current main	This PR
k=10, 250 queries	120 mismatches	0 mismatches
k=100, 250 queries	1 mismatch	0 mismatches
Total	121 / 500 mismatches	500 / 500 exact

The target also had zero row/result signature instability across four timed processes.

This is primarily a correctness fix. Current main's higher throughput is not a valid performance baseline because its result set is incorrect on the affected queries.

Tests and checks

• Added explicit and fieldless top-level cross-column MultiMatch oracle coverage.
• Covers equal-score kth ties, reversed segment order, signed zero boosts, unbounded execution, and mixed bounded/exact planning with one partially covered field.
• cargo check -p lance --tests
• cargo fmt --all -- --check
• git diff --check
• Pre-commit fmt and typos checks

Per the requested workflow, I did not run local cargo test or cargo clippy; CI will run them.

[lance-gatekeeper]

✅ Gate recommendation: approve.

Both prior findings are fixed. The revised planner preserves each eligible field’s exact bounded top-k scorer, confines exhaustive fallback to ineligible fields, and the mixed partial-coverage result matches the exhaustive oracle.