perf: add parallel implementation to the numerical_utils module's multiply_columns function (#1002)

# Rationale for this change
This PR adds a parallel implementation to the `numerical_utils` module's
`multiply_columns` function. Various benchmarks have improved with this
change. For example, the Sum Count benchmark query had a 1.42x
performance improvement on the Multi-A100 VM with a table size of
`1,000,000`. The `MultiplyExpr::final_round_evaluate` saw a 9.42x
performance improvement.

Before 2.39s
<img width="1878" height="341" alt="image"
src="https://github.com/user-attachments/assets/b9305381-d162-4f2d-8024-221d96e4c0b0"
/>

After 1.72s
<img width="1878" height="341" alt="image"
src="https://github.com/user-attachments/assets/5c128848-685d-494d-a0ff-bff30988ab91"
/>

This is a first step. Further improvements will investigate doing
multiplication inside of the Column module to avoid conversions during
each multiplication.

Note: creating an `unsafe_scalar_at` function that avoid returning an
Option did not show any performance gains when benchmarked.

# What changes are included in this PR?
- A Rayon implementation is added to the the `numerical_utils` module's
`multiply_columns` function

# Are these changes tested?
Yes

Author: jacobtrombetta

crates/proof-of-sql/src/sql/proof_exprs/numerical_util.rs

@@ -1,5 +1,6 @@
 use crate::base::{
     database::{try_cast_types, try_scale_cast_types, Column, ColumnOperationResult, ColumnType},
+    if_rayon,
     math::decimal::Precision,
     scalar::{Scalar, ScalarExt},
 };
@@ -9,6 +10,8 @@ use bumpalo::Bump;
 use core::{convert::TryInto, ops::Neg};
 use itertools::izip;
 use num_traits::{NumCast, PrimInt};
+#[cfg(feature = "rayon")]
+use rayon::iter::{IndexedParallelIterator, IntoParallelRefMutIterator, ParallelIterator};
 
 /// Add or subtract two columns together.
 #[tracing::instrument(level = "debug", skip_all)]
@@ -51,9 +54,17 @@ pub(crate) fn multiply_columns<'a, S: Scalar>(
         lhs_len == rhs_len,
         "lhs and rhs should have the same length"
     );
-    alloc.alloc_slice_fill_with(lhs_len, |i| {
-        lhs.scalar_at(i).unwrap() * rhs.scalar_at(i).unwrap()
-    })
+    if_rayon!(
+        {
+            let result = alloc.alloc_slice_fill_with(lhs_len, |_| S::ZERO);
+            result.par_iter_mut().enumerate().for_each(|(i, val)| {
+                *val = lhs.scalar_at(i).unwrap() * rhs.scalar_at(i).unwrap();
+            });
+            result
+        },
+        alloc.alloc_slice_fill_with(lhs_len, |i| lhs.scalar_at(i).unwrap()
+            * rhs.scalar_at(i).unwrap())
+    )
 }
 
 /// Divides two columns of data, where the data types are some signed int type(s).