fix(index): correct cosine RabitQ query factor

[lance-gatefixer]

Summary

• treat normalized cosine as L2 when calculating the residual-query RabitQ partition factor
• preserve the Dot-specific offset
• compare full cosine and L2 calculator outputs in a regression test

Root cause

Cosine vectors are normalized before IVF_RQ search, so the residual-query estimator operates on squared L2 distances. Its query-factor match still grouped Cosine with Dot and subtracted 1.0, shifting every approximate cosine distance.

Validation

• cargo test -p lance-index test_residual_query_cosine_uses_l2_query_factor
• cargo fmt --all -- --check
• cargo clippy --all --tests --benches -- -D warnings

Fixes #5649

[lance-gatekeeper]

✅ Gate recommendation: approve.

The residual-query estimator now uses the squared-L2 partition factor required by normalized cosine vectors while preserving the distinct dot-product offset. This is the narrow fix for the incorrect legacy score and bound semantics, and the focused regression covers the affected calculator path.

[cursor]

Review: LGTM — approve & merge

For ResidualQuery, Cosine already works on normalized L2 residuals, so q_factor should match L2 (dist_q_c) rather than Dot (dist_q_c - 1.0). Focused change; test now asserts full distance_all equality between Cosine and L2 residual paths.