…nsport (-20%) (#10837)

Switch the JS->Rust raw query results wire format from row-oriented `Vec<IndexMap<String, DBResponsePrimitive>>` (each row carrying its own copy of every column name) to columnar `JsRawColumnarData { members, columns }`. The pivot is done once on the JS side in `ResultWrapper.serialize` before `JSON.stringify`, then `from_js_raw_data` transposes back to the existing row-major `QueryResult` on Rust without per-cell clones.

Also drops the `JsArray` + `JsValueDeserializer` ingest path in `orchestrator.rs` — only `JsBox<Arc<QueryResult>>` and `JsBuffer` are accepted now. The `JsValueDeserializer` walk was the slow path the buffer route was already winning against.

Bench: synthetic dataset, 60% string dims / 40% f64 measures, no nulls,
`cargo bench --bench transform -- --sample-size 30 --measurement-time 3`.

Rust ingest path (`serde_json::from_slice` + `from_js_raw_data`, median):

| Dataset             | rows (master) | columnar (this) | Faster by | Speedup |
| ------------------- | ------------: | --------------: | --------: | ------: |
| 8 cols x 10k rows   |       52.4 ms |         40.8 ms |    -22 %  |  1.29 x |
| 16 cols x 10k rows  |      107.2 ms |         82.9 ms |    -23 %  |  1.29 x |
| 16 cols x 100k rows |     1024.7 ms |        827.7 ms |    -19 %  |  1.24 x |
| 32 cols x 100k rows |     2019.8 ms |       1633.9 ms |    -19 %  |  1.24 x |