fix(index): stop appending a spurious byte in binary_quantization

[LuciferYang]

Problem

binary_quantization appends the tail byte unconditionally, so an input whose length is a multiple of 8 produces floor(len / 8) + 1 bytes instead of ceil(len / 8). Through BinaryQuantization::transform that is one extra zero byte per row for any dimension that is a multiple of 8.

Fixes #9028.

What this changes

Compute the tail byte once and chain it only when chunks_exact(8) left a remainder. The closure mapped over the full chunks is untouched, so the loop the "auto vectorized" comment refers to is the same code as before.

BinaryQuantization has no callers in this repository, so no in-tree width changes. The old width also contradicted the crate's own sizing helper, rabit_binary_code_bytes, which is div_ceil(8). Deleting the type is the other option, since RabitQuantizer is what IVF_RQ uses, but it is public API and the behaviour is cheap to fix.

Test plan

test_binary_quantization_length_multiple_of_eight pins the helper at 16 values, and two tests pin the public entry point on both branches: 3 rows of dimension 8, and 2 rows of dimension 12 where the tail byte must still appear. Restoring the old implementation makes them fail with [85, 85, 0] against [85, 85] and 6 bytes against 3. The existing test_binary_quantization covers the remainder case at 11 values, so both branches stay asserted.

• cargo test -p lance-index --lib vector::bq 233 passed
• cargo clippy --all --tests --benches -- -D warnings clean
• cargo fmt --all --check clean

[lance-gatekeeper]

✅ Gate recommendation: approve.

The conditional tail emission fixes the boundary error at its root while preserving bit ordering and the performance-sensitive full-chunk loop.