perf/sorter: sort pointers instead of records, use arena allocation

1. Use bump arena for in-memory records
- Allows contiguous allocation and bulk deallocate
- Allows sorting pointers instead of sorting large SortableImmutableRecord structs which pay a large memmove penalty
2. Use boxed records for on-disk chunk heap: avoids the same memmove problem

Author: jussisaurio

Cargo.lock

@@ -88,6 +88,7 @@ rustc-hash = "2.0"
 either = { workspace = true }
 tracing-subscriber.workspace = true
 rapidhash = "4.1.1"
+bumpalo = { version = "3", features = ["collections"] }
 
 # Use pure-rust for Android to avoid C cross-compilation issues
 [target.'cfg(target_os = "android")'.dependencies]

core/Cargo.toml

@@ -1311,7 +1311,12 @@ impl RecordCursor {
         record: &ImmutableRecord,
         target_idx: usize,
     ) -> Result<()> {
-        let payload = record.get_payload();
+        self.ensure_parsed_upto_payload(record.get_payload(), target_idx)
+    }
+
+    /// Like `ensure_parsed_upto` but works directly with raw payload bytes.
+    #[inline(always)]
+    pub fn ensure_parsed_upto_payload(&mut self, payload: &[u8], target_idx: usize) -> Result<()> {
         if payload.is_empty() {
             return Ok(());
         }
@@ -1365,6 +1370,15 @@ impl RecordCursor {
         &self,
         record: &'a ImmutableRecord,
         idx: usize,
+    ) -> Result<ValueRef<'a>> {
+        self.deserialize_column_payload(record.get_payload(), idx)
+    }
+
+    /// Like `deserialize_column` but works directly with raw payload bytes.
+    pub fn deserialize_column_payload<'a>(
+        &self,
+        payload: &'a [u8],
+        idx: usize,
     ) -> Result<ValueRef<'a>> {
         if idx >= self.serial_types.len() {
             return Ok(ValueRef::Null);
@@ -1386,7 +1400,6 @@ impl RecordCursor {
 
         let start = self.offsets[idx];
         let end = self.offsets[idx + 1];
-        let payload = record.get_payload();
 
         let slice = &payload[start..end];
         let (value, _) = crate::storage::sqlite3_ondisk::read_value(slice, serial_type_obj)?;