fix(index): store the converted metric in IvfTransformer::new_flat

[LuciferYang]

Closes #9398

new_flat converts cosine to L2 for its transform chain (normalize, then assign with L2) and then handed the original metric to IvfTransformer::new, unlike with_pq and the constructor above it. The stored metric is what the transformer's own find_partitions, compute_partitions, and compute_residual use, so a cosine flat transformer measured in cosine while its chain had assigned in L2. It now stores dt, so all four agree.

Partition choice is unchanged: the vectors and centroids are normalized by then, and cosine and L2 rank unit vectors the same way. What changes is the scale of the distances those three methods return, which now matches the __ivf_centroid_dist values the chain produces.

How was this patch tested?

test_new_flat_keeps_the_converted_metric builds a cosine and an L2 flat transformer over the same unit centroids and asserts find_partitions returns the same partitions and the same distances for both. With the original metric stored, the cosine one returns cosine distances and the test fails. The file had no test module before, so this adds one.

[lance-gatekeeper]

✅ Gate recommendation: approve.

Storing the converted L2 metric matches the transformer’s existing normalize-then-L2 contract and prevents the Cosine variant from reaching unsupported partition-distance code. The added regression covers this invariant; I found no blocking or material residual risk.