fix(index): return zero balance factor when no vector is assigned

[LuciferYang]

Closes #9406

compute_cluster_sizes divided the largest cluster's loss by that cluster's size, which is zero when every membership is None, so it returned NaN. The caller's adjusted_balance_factor.min(params.balance_factor) then dropped the NaN because that is what f32::min does, leaving the next iteration's factor decided by an accident rather than by the code. It now returns 0.0 for that case, which says what the situation is: nothing was assigned, so there is nothing to balance.

The only inputs that reach it are batches where every vector is non-finite, which the index build path filters out before training; direct KMeans users can still hit it.

How was this patch tested?

test_compute_cluster_sizes_without_any_assignment calls the function with an all-None membership and asserts the factor is 0.0 and finite, and that the cluster sizes were reset. Removing the guard returns NaN and fails the assertion.

[lance-gatekeeper]

✅ Gate recommendation: approve.

The zero-assignment guard uses the documented 0.0 “no balancing” value, keeps NaN out of the next iteration, and preserves the cluster-size reset. The focused regression test covers the all-None boundary.

[LuciferYang]

The rustdoc job is failing on main as well, not because of this change: run 35372502724 on main (83d0084) fails with 'error: could not document lance' from rust/lance/src/io/exec/knn.rs:563, a private intra-doc link to Self::try_new_batch under -D rustdoc::private-intra-doc-links. This PR touches rust/lance-index/src/vector/kmeans.rs only.