ScanIterByColRange: Avoid unnecessary ProductValue allocations (#804)

Prior to this commit, `ScanIterByColRange::next` called `RowRef::to_product_value`
and then `ProductValue::project_not_empty` on every row in the table
to get a key, which it compared to the sought range.
This always allocated at least a `ProductValue`,
even when seeking a range of primitive type like `u64`.
It also unnecessarily deserialized (and potentially allocated)
columns not relevant to the search for rows which did not match the range.

With this commit, we call `RowRef::project_not_empty` directly.
When the sought range is a single column, this avoids allocating a `ProductValue`.
Even when seeking a multi-column range, this avoids deserializing unrelated columns
of non-matching rows.

Non-indexed `IterByColEq` is always going to be slow, so this probably doesn't matter,
but it's a trivial change.

Author: gefjon

crates/core/src/db/datastore/locking_tx_datastore/state_view.rs

@@ -490,8 +490,7 @@ impl<'a, R: RangeBounds<AlgebraicValue>> Iterator for ScanIterByColRange<'a, R>
     #[tracing::instrument(skip_all)]
     fn next(&mut self) -> Option<Self::Item> {
         for row_ref in &mut self.scan_iter {
-            let row = row_ref.to_product_value();
-            let value = row.project_not_empty(&self.cols).unwrap();
+            let value = row_ref.project_not_empty(&self.cols).unwrap();
             if self.range.contains(&value) {
                 return Some(row_ref);
             }